Point cloud data encoding device, point cloud data encoding method, point cloud data decoding device, and point cloud data decoding method

Abstract

A method according to embodiments may comprise the steps of: decoding geometry data of point cloud data within a bitstream; and decoding attribute data of the point cloud data. A method according to embodiments may comprise the steps of: encoding geometry data of point cloud data; and encoding attribute data of the point cloud data.

Description

Point cloud data encoding device, point cloud data encoding method, point cloud data decoding device and point cloud data decoding method

[0001] The embodiments relate to a method and apparatus for processing point cloud content.

[0002] Point cloud content is content represented as a point cloud, which is a set of points belonging to a coordinate system that represents three-dimensional space. Point cloud content can represent three-dimensional media and is used to provide various services such as VR (Virtual Reality), AR (Augmented Reality), MR (Mixed Reality), and autonomous driving services. However, representing point cloud content requires tens of thousands to hundreds of thousands of point data points. Therefore, a method is required to efficiently process a vast amount of point data.

[0003] The embodiments provide an apparatus and a method for efficiently processing point cloud data. The embodiments provide a method and apparatus for processing point cloud data to address latency and encoding / decoding complexity.

[0004] However, the scope of rights of the embodiments is not limited to the technical problems described above, and may be extended to other technical problems that a person skilled in the art can infer based on the entire content described.

[0005] A method according to the embodiments may include the step of decoding geometry data of point cloud data within a bitstream; and the step of decoding attribute data of point cloud data. A method according to the embodiments may include the step of encoding geometry data of point cloud data; and the step of encoding attribute data of point cloud data.

[0006] The device and method according to the embodiments can process point cloud data with high efficiency.

[0007] The device and method according to the embodiments can provide a high-quality point cloud service.

[0008] The device and method according to the embodiments can provide point cloud content for providing general-purpose services such as VR services and autonomous driving services.

[0009] Drawings are included to further understand the embodiments, and the drawings illustrate the embodiments along with descriptions related to the embodiments. For a better understanding of the various embodiments described below, one must refer to the description of the embodiments below in relation to the following drawings, which include parts corresponding to similar reference numerals throughout the drawings.

[0010] FIG. 1 shows an example of a point cloud content provision system according to embodiments.

[0011] FIG. 2 is a block diagram illustrating a point cloud content provision operation according to embodiments.

[0012] FIG. 3 shows an example of a point cloud encoder according to embodiments.

[0013] FIG. 4 shows examples of octree and occupancy codes according to embodiments.

[0014] Figure 5 shows an example of a point configuration by LOD according to embodiments.

[0015] Figure 6 shows an example of a point configuration by LOD according to embodiments.

[0016] FIG. 7 shows an example of a point cloud decoder according to embodiments.

[0017] FIG. 8 is an example of a transmission device according to embodiments.

[0018] FIG. 9 is an example of a receiving device according to embodiments.

[0019] FIG. 10 shows an example of a structure that can be linked with a point cloud data transmission / reception method / device according to embodiments.

[0020] FIG. 11 is an example of point cloud segmentation according to embodiments.

[0021] FIG. 12 shows an example in which an original signal and a reconstructed signal according to embodiments are contained in a single cube.

[0022] FIG. 13 shows an example in which two different point clouds of the same size and two spheres with radii of 16 and 2 are added for comparison according to embodiments.

[0023] FIG. 14 shows examples of Z-order curves (Lebesgue curve 1904, Morton space-filling curve 1966) according to embodiments.

[0024] FIG. 15 shows an example of a 2D projection of a block defined by 8 corner phoxels and 27 corners + middle corners + center phoxels according to embodiments.

[0025] FIG. 16 shows an example of retrieving truncated_block_code from a Morton address of a voxel according to embodiments.

[0026] FIG. 17 shows information about a Molten code for neighbor search according to embodiments.

[0027] FIGS. 18, 19, and 20 show filter window sizes according to embodiments.

[0028] FIG. 21 shows the ploxel distribution of filter tabs according to distance (2D and 3D cases) according to embodiments.

[0029] FIG. 22 shows the shape of a filter according to the embodiments.

[0030] FIG. 23 shows the phoxel distribution (3D case) of filter tabs according to the direction according to the embodiments.

[0031] FIG. 24 shows an example of a 4-tab symmetric ALF design using 8 octants according to embodiments.

[0032] FIG. 25 shows a flowchart of scaling and rounding according to embodiments.

[0033] FIG. 26 shows a flowchart of the calculation of ALF tap values ​​according to embodiments.

[0034] FIG. 27 shows the dynamic range for calculating a four-tap filter according to the embodiments.

[0035] FIG. 28 illustrates an RDO-based application of ALF for various QPs according to embodiments.

[0036] FIG. 29 shows an example of the most interesting part in the image according to the embodiments.

[0037] FIG. 30 shows the APS (attribute parameter set) syntax in the bitstream.

[0038] FIG. 31 shows an attribute data unit according to embodiments.

[0039] FIG. 32 shows an attribute data unit header according to embodiments.

[0040] FIG. 33 shows a flowchart of ALF coefficient coding according to embodiments.

[0041] FIG. 34 shows a flowchart of ALF coefficient coding for four tap filters encodeCoeff, encodeCoeffLast, and encodeCoeffzeros according to embodiments.

[0042] FIG. 35 shows a flowchart of an ALF coefficient decoding process according to embodiments.

[0043] FIGS. 36a and FIGS. 36b show attribute ALF coefficients according to the embodiments.

[0044] FIG. 37 shows an attribute ALF coefficient tuple according to the embodiments.

[0045] FIG. 38 shows an encoding device according to embodiments.

[0046] FIG. 39 shows a decoding device according to embodiments.

[0047] FIG. 40 illustrates a encoding method according to embodiments.

[0048] FIG. 41 illustrates a decoding method according to embodiments.

[0049] Preferred embodiments of the embodiments are described in detail, and examples thereof are shown in the accompanying drawings. The following detailed description, with reference to the accompanying drawings, is intended to describe preferred embodiments of the embodiments rather than merely embodiments that may be implemented according to the embodiments. The following detailed description includes details to provide a thorough understanding of the embodiments. However, it is obvious to those skilled in the art that the embodiments may be practiced without these details.

[0050] Most terms used in the embodiments are selected from those commonly used in the field, but some terms are chosen at the applicant's discretion, and their meanings are described in detail in the following description as necessary. Accordingly, the embodiments should be understood based on the intended meaning of the terms, rather than their mere names or meanings.

[0051] FIG. 1 shows an example of a point cloud content provision system according to embodiments.

[0052] The point cloud content providing system illustrated in FIG. 1 may include a transmission device (10000) and a reception device (10004). The transmission device (10000) and the reception device (10004) can communicate via wired or wireless means to transmit and receive point cloud data.

[0053] A transmission device (10000) according to embodiments can acquire, process, and transmit point cloud video (or point cloud content). According to embodiments, the transmission device (10000) may include a fixed station, a base transceiver system (BTS), a network, an Artificial Intelligence (AI) device and / or system, a robot, an AR / VR / XR device and / or server, etc. Additionally, according to embodiments, the transmission device (10000) may include a device that communicates with a base station and / or other wireless devices using wireless access technology (e.g., 5G NR (New RAT), LTE (Long Term Evolution)), a robot, a vehicle, an AR / VR / XR device, a mobile device, a home appliance, an Internet of Things (IoT) device, an AI device / server, etc.

[0054] A transmission device (10000) according to embodiments includes a point cloud video acquisition unit (10001), a point cloud video encoder (10002), and / or a transmitter (or communication module), 10003.

[0055] A point cloud video acquisition unit (10001) according to the embodiments acquires a point cloud video through processing steps such as capture, synthesis, or generation. The point cloud video is a point cloud content represented as a point cloud, which is a set of points located in a three-dimensional space, and may be referred to as point cloud video data, etc. The point cloud video according to the embodiments may include one or more frames. A frame represents a still image / picture. Accordingly, the point cloud video may include a point cloud image / frame / picture and may be referred to as any one of a point cloud image, a frame, and a picture.

[0056] A point cloud video encoder (10002) according to the embodiments encodes the obtained point cloud video data. The point cloud video encoder (10002) can encode the point cloud video data based on point cloud compression coding. The point cloud compression coding according to the embodiments may include Geometry-based Point Cloud Compression (G-PCC) coding and / or Video-based Point Cloud Compression (V-PCC) coding or next-generation coding. Furthermore, the point cloud compression coding according to the embodiments is not limited to the embodiments described above. The point cloud video encoder (10002) can output a bitstream containing the encoded point cloud video data. The bitstream may include not only the encoded point cloud video data but also signaling information related to the encoding of the point cloud video data.

[0057] A transmitter (10003) according to the embodiments transmits a bitstream containing encoded point cloud video data. The bitstream according to the embodiments is encapsulated into a file or segment (e.g., a streaming segment) and transmitted through various networks such as a broadcast network and / or a broadband network. Although not illustrated in the drawings, the transmission device (10000) may include an encapsulation unit (or encapsulation module) that performs an encapsulation operation. Additionally, according to the embodiments, the encapsulation unit may be included in the transmitter (10003). According to the embodiments, the file or segment may be transmitted to a receiving device (10004) via a network or stored on a digital storage medium (e.g., USB, SD, CD, DVD, Blu-ray, HDD, SSD, etc.). The transmitter (10003) according to the embodiments can communicate wired or wirelessly with the receiving device (10004) (or receiver (10005)) via a network such as 4G, 5G, or 6G. Additionally, the transmitter (10003) can perform necessary data processing operations according to a network system (e.g., a communication network system such as 4G, 5G, 6G, etc.). Additionally, the transmission device (10000) can transmit encapsulated data according to an on-demand method.

[0058] A receiving device (10004) according to embodiments includes a receiver (10005), a point cloud video decoder (10006), and / or a renderer (10007). According to embodiments, the receiving device (10004) may include a device, robot, vehicle, AR / VR / XR device, mobile device, home appliance, IoT (Internet of Thing) device, AI device / server, etc., that communicates with a base station and / or other wireless device using wireless access technology (e.g., 5G NR (New RAT), LTE (Long Term Evolution)).

[0059] A receiver (10005) according to the embodiments receives a bitstream containing point cloud video data or a file / segment containing the bitstream from a network or a storage medium. The receiver (10005) can perform necessary data processing operations according to a network system (e.g., a communication network system such as 4G, 5G, 6G, etc.). The receiver (10005) according to the embodiments can output a bitstream by decapsulating the received file / segment. Additionally, according to the embodiments, the receiver (10005) may include a decapsulation unit (or decapsulation module) for performing a decapsulation operation. Additionally, the decapsulation unit may be implemented as an element (or component) separate from the receiver (10005).

[0060] A point cloud video decoder (10006) decodes a bitstream containing point cloud video data. The point cloud video decoder (10006) can decode the point cloud video data according to the way the point cloud video data is encoded (e.g., the reverse process of the operation of a point cloud video encoder (10002)). Accordingly, the point cloud video decoder (10006) can decode the point cloud video data by performing point cloud decompression coding, which is the reverse process of point cloud compression. Point cloud decompression coding includes G-PCC coding.

[0061] The renderer (10007) renders the decoded point cloud video data. The renderer (10007) can render not only the point cloud video data but also audio data to output point cloud content. According to embodiments, the renderer (10007) may include a display for displaying the point cloud content. According to embodiments, the display may not be included in the renderer (10007) but may be implemented as a separate device or component.

[0062] The arrows indicated by dotted lines in the drawing represent the transmission path of feedback information obtained from the receiving device (10004). The feedback information is information intended to reflect interaction with a user consuming point cloud content, and includes user information (e.g., head orientation information), viewport information, etc. In particular, if the point cloud content is content for a service requiring interaction with a user (e.g., autonomous driving service, etc.), the feedback information may be transmitted to the content transmission side (e.g., transmission device (10000)) and / or the service provider. Depending on the embodiments, the feedback information may be used in the receiving device (10004) as well as the transmission device (10000), or it may not be provided.

[0063] Head orientation information according to the embodiments is information regarding the user's head position, direction, angle, movement, etc. The receiving device (10004) according to the embodiments can calculate viewport information based on the head orientation information. Viewport information is information about the area of ​​the point cloud video that the user is looking at. The viewpoint refers to the point where the user is looking at the point cloud video, and may mean the exact center point of the viewport area. That is, the viewport is an area centered on the viewpoint, and the size and shape of the area can be determined by the Field Of View (FOV). Therefore, the receiving device (10004) can extract viewport information based on the vertical or horizontal FOV supported by the device in addition to the head orientation information. In addition, the receiving device (10004) performs gaze analysis, etc., to check the user's point cloud consumption method, the point cloud video area the user is looking at, the gaze time, etc. According to embodiments, the receiving device (10004) may transmit feedback information including gaze analysis results to the transmitting device (10000). According to embodiments, the feedback information may be obtained during the rendering and / or display process. According to embodiments, the feedback information may be obtained by one or more sensors included in the receiving device (10004). Also, according to embodiments, the feedback information may be obtained by the renderer (10007) or a separate external element (or device, component, etc.). The dotted line in FIG. 1 indicates the process of transmitting the feedback information obtained from the renderer (10007). The point cloud content providing system may process (encode / decode) point cloud data based on the feedback information. Accordingly, the point cloud video data decoder (10006) may perform a decoding operation based on the feedback information.Additionally, the receiving device (10004) can transmit feedback information to the transmitting device (10000). The transmitting device (10000) (or the point cloud video data encoder (10002)) can perform an encoding operation based on the feedback information. Thus, the point cloud content providing system can efficiently process necessary data (e.g., point cloud data corresponding to the user's head position) based on the feedback information without processing (encoding / decoding) all point cloud data, and provide point cloud content to the user.

[0064] According to embodiments, the transmission device (10000) may be referred to as an encoder, transmission device, transmitter, etc., and the receiving device (10004) may be referred to as a decoder, receiving device, receiver, etc.

[0065] Point cloud data processed in the point cloud content providing system of FIG. 1 according to embodiments (processed through a series of processes of acquisition / encoding / transmission / decoding / rendering) may be referred to as point cloud content data or point cloud video data. According to embodiments, point cloud content data may be used as a concept including metadata or signaling information related to point cloud data.

[0066] The elements of the point cloud content delivery system illustrated in FIG. 1 can be implemented in hardware, software, processors, and / or combinations thereof.

[0067] FIG. 2 is a block diagram illustrating a point cloud content provision operation according to embodiments.

[0068] The block diagram of FIG. 2 illustrates the operation of the point cloud content provision system described in FIG. 1. As described above, the point cloud content provision system can process point cloud data based on point cloud compression coding (e.g., G-PCC).

[0069] A point cloud content providing system according to the embodiments (e.g., a point cloud transmission device (10000) or a point cloud video acquisition unit (10001)) can acquire a point cloud video (20000). The point cloud video is represented as a point cloud belonging to a coordinate system representing a three-dimensional space. The point cloud video according to the embodiments may include a Ply (Polygon File format or the Stanford Triangle format) file. If the point cloud video has one or more frames, the acquired point cloud video may include one or more Ply files. The Ply file contains point cloud data such as the geometry and / or attributes of the points. The geometry includes the positions of the points. The position of each point may be represented by parameters (e.g., values ​​of the X-axis, Y-axis, and Z-axis, respectively) representing a three-dimensional coordinate system (e.g., a coordinate system consisting of XYZ axes). Attributes include attributes of points (e.g., texture information, color (YCbCr or RGB), reflectance (r), transparency, etc. of each point). A point has one or more attributes (or properties). For example, a point may have one attribute which is color, or two attributes which are color and reflectance. According to embodiments, geometry may be referred to as positions, geometry information, geometry data, etc., and attributes may be referred to as attributes, attribute information, attribute data, etc.In addition, a point cloud content provision system (e.g., a point cloud transmission device (10000) or a point cloud video acquisition unit (10001)) can obtain point cloud data from information related to the acquisition process of point cloud video (e.g., depth information, color information, etc.).

[0070] A point cloud content providing system according to embodiments (e.g., a transmission device (10000) or a point cloud video encoder (10002)) can encode point cloud data (20001). The point cloud content providing system can encode point cloud data based on point cloud compression coding. As described above, point cloud data may include geometry and attributes of points. Accordingly, the point cloud content providing system can output a geometry bitstream by performing geometry encoding to encode geometry. The point cloud content providing system can output an attribute bitstream by performing attribute encoding to encode attributes. According to embodiments, the point cloud content providing system can perform attribute encoding based on geometry encoding. The geometry bitstream and attribute bitstream according to embodiments can be multiplexed and output as a single bitstream. The bitstream according to the embodiments may further include signaling information related to geometry encoding and attribute encoding.

[0071] A point cloud content providing system according to embodiments (e.g., a transmission device (10000) or a transmitter (10003)) can transmit encoded point cloud data (20002). As described in FIG. 1, the encoded point cloud data can be represented as a geometry bitstream and an attribute bitstream. Additionally, the encoded point cloud data can be transmitted in the form of a bitstream along with signaling information related to the encoding of the point cloud data (e.g., signaling information related to geometry encoding and attribute encoding). Additionally, the point cloud content providing system can encapsulate the bitstream transmitting the encoded point cloud data and transmit it in the form of a file or segment.

[0072] A point cloud content providing system according to embodiments (e.g., a receiving device (10004) or a receiver (10005)) can receive a bitstream containing encoded point cloud data. Additionally, the point cloud content providing system (e.g., a receiving device (10004) or a receiver (10005)) can demultiplex the bitstream.

[0073] A point cloud content providing system (e.g., a receiving device (10004) or a point cloud video decoder (10005)) can decode encoded point cloud data (e.g., a geometry bitstream, an attribute bitstream) transmitted as a bitstream. A point cloud content providing system (e.g., a receiving device (10004) or a point cloud video decoder (10005)) can decode point cloud video data based on signaling information related to the encoding of point cloud video data included in the bitstream. A point cloud content providing system (e.g., a receiving device (10004) or a point cloud video decoder (10005)) can decode the geometry bitstream to restore the positions (geometry) of the points. A point cloud content providing system can decode the attribute bitstream based on the restored geometry to restore the attributes of the points. A point cloud content delivery system (e.g., a receiving device (10004) or a point cloud video decoder (10005)) can restore a point cloud video based on positions according to the restored geometry and decoded attributes.

[0074] A point cloud content providing system according to embodiments (e.g., a receiving device (10004) or a renderer (10007)) can render decoded point cloud data (20004). The point cloud content providing system (e.g., a receiving device (10004) or a renderer (10007)) can render geometry and attributes decoded through a decoding process according to various rendering methods. Points of the point cloud content may be rendered as vertices having a certain thickness, cubes having a specific minimum size with the vertex location as the center, or circles with the vertex location as the center, etc. All or part of the rendered point cloud content is provided to a user through a display (e.g., a VR / AR display, a general display, etc.).

[0075] A point cloud content providing system (e.g., a receiving device (10004)) according to the embodiments can obtain feedback information (20005). The point cloud content providing system can encode and / or decode point cloud data based on the feedback information. Since the feedback information and the operation of the point cloud content providing system according to the embodiments are the same as the feedback information and operation described in FIG. 1, a detailed description is omitted.

[0076] FIG. 3 shows an example of a point cloud encoder according to embodiments.

[0077] FIG. 3 shows an example of the point cloud video encoder (10002) of FIG. 1. The point cloud encoder reconstructs point cloud data (e.g., positions and / or attributes of points) and performs encoding operations to adjust the quality of point cloud content (e.g., lossless, lossy, near-lossless) according to network conditions or applications. If the total size of the point cloud content is large (e.g., point cloud content of 60 Gbps in the case of 30 fps), the point cloud content delivery system may not be able to stream the content in real time. Therefore, the point cloud content delivery system may reconstruct the point cloud content based on a maximum target bitrate to provide it according to the network environment.

[0078] As described in FIGS. 1 and 2, the point cloud encoder can perform geometry encoding and attribute encoding. Geometry encoding is performed before attribute encoding.

[0079] The point cloud encoder according to the embodiments comprises a coordinate system transformation unit (Transformation Coordinates, 30000), a quantization unit (Quantize and Remove Points (Voxelize), 30001), an octree analysis unit (Analyze Octree, 30002), a surface approximation analysis unit (Analyze Surface Approximation, 30003), an arithmetic encoder (Arithmetic Encode, 30004), a geometry reconstruction unit (Reconstruct Geometry, 30005), a color transformation unit (Transform Colors, 30006), an attribute transformation unit (Transfer Attributes, 30007), a RAHT transformation unit (30008), an LOD generation unit (Generated LOD, 30009), a lifting transformation unit (Lifting) (30010), and a coefficient quantization unit (Quantize Coefficients, 30011). Includes an and / or arithmetic encoder (Arithmetic Encode, 30012).

[0080] The coordinate system transformation unit (30000), quantization unit (30001), octree analysis unit (30002), surface approximation analysis unit (30003), arismetic encoder (30004), and geometry reconstruction unit (30005) can perform geometry encoding. Geometry encoding according to the embodiments may include octree geometry coding, direct coding, trisoup geometry encoding, and entropy encoding. Direct coding and trisoup geometry encoding are applied optionally or in combination. Additionally, geometry encoding is not limited to the above examples.

[0081] As illustrated in the drawings, the coordinate system conversion unit (30000) according to the embodiments receives positions and converts them into a coordinate system. For example, the positions can be converted into position information in a three-dimensional space (e.g., a three-dimensional space expressed in an XYZ coordinate system). The position information in the three-dimensional space according to the embodiments may be referred to as geometry information.

[0082] The quantization unit (30001) according to the embodiments quantizes the geometry. For example, the quantization unit (30001) can quantize points based on the minimum position values ​​of all points (e.g., minimum values ​​on each axis for the X-axis, Y-axis, and Z-axis). The quantization unit (30001) performs a quantization operation to find the nearest integer value by multiplying the difference between the minimum position value and the position value of each point by a preset quantization scale value and then performing rounding down or rounding up. Thus, one or more points may have the same quantized position (or position value). The quantization unit (30001) according to the embodiments performs voxelization based on the quantized positions to reconstruct the quantized points. Just as the minimum unit containing 2D image / video information is a pixel, the points of the point cloud content (or 3D point cloud video) according to the embodiments may be contained in one or more voxels. A voxel is a combination of volume and pixel, and refers to a three-dimensional cubic space that is generated when a three-dimensional space is divided into units (unit=1.0) based on axes representing the three-dimensional space (e.g., X-axis, Y-axis, Z-axis). The quantization unit (40001) can match groups of points in the three-dimensional space to voxels. According to embodiments, a single voxel may contain only one point. According to embodiments, a single voxel may contain one or more points. In addition, to represent a single voxel as a single point, the position of the center of the voxel can be set based on the positions of one or more points included in the voxel. In this case, the attributes of all positions included in the voxel can be combined and assigned to the voxel.

[0083] The octree analysis unit (30002) according to the embodiments performs octree geometry coding (or octree coding) to represent the voxels in an octree structure. The octree structure represents points matched to the voxels based on an octree structure.

[0084] The surface approximation analysis unit (30003) according to the embodiments can analyze and approximate an octree. The octree analysis and approximation according to the embodiments is a process of analyzing to voxelize an area containing multiple points in order to efficiently provide octree and voxelization.

[0085] An arithmetic encoder (30004) according to the embodiments entropy-encodes an octree and / or an approximated octree. For example, the encoding method includes an arithmetic encoding method. As a result of the encoding, a geometry bitstream is generated.

[0086] The color conversion unit (30006), attribute conversion unit (30007), RAHT conversion unit (30008), LOD generation unit (30009), lifting conversion unit (30010), coefficient quantization unit (30011) and / or arismetic encoder (30012) perform attribute encoding. As described above, a point may have one or more attributes. The attribute encoding according to the embodiments is applied equally to the attributes of a point. However, if a single attribute (e.g., color) includes one or more elements, independent attribute encoding is applied to each element. The attribute encoding according to the embodiments may include color conversion coding, attribute conversion coding, Region Adaptive Hierarchial Transform (RAHT) coding, prediction transformation (Interpolaration-based hierarchical nearest-neighbour prediction-Prediction Transform) coding, and lifting transformation (interpolation-based hierarchical nearest-neighbour prediction with an update / lifting step (Lifting Transform)) coding. Depending on the point cloud content, the above-described RAHT coding, prediction transformation coding, and lifting transformation coding may be used optionally, or a combination of one or more of the codings may be used. Furthermore, the attribute encoding according to the embodiments is not limited to the examples described above.

[0087] The color conversion unit (30006) according to the embodiments performs color conversion coding that converts color values ​​(or textures) included in attributes. For example, the color conversion unit (30006) can convert the format of color information (e.g., convert from RGB to YCbCr). The operation of the color conversion unit (30006) according to the embodiments may be applied optionally depending on the color values ​​included in attributes.

[0088] The geometry reconstruction unit (30005) according to the embodiments reconstructs (decompresses) an octree and / or an approximated octree. The geometry reconstruction unit (30005) reconstructs an octree / voxel based on the results of analyzing the distribution of points. The reconstructed octree / voxel may be referred to as the reconstructed geometry (or restored geometry).

[0089] The attribute transformation unit (30007) according to the embodiments performs attribute transformation that transforms attributes based on positions where geometry encoding has not been performed and / or reconstructed geometry. As described above, since attributes are dependent on geometry, the attribute transformation unit (30007) can transform attributes based on reconstructed geometry information. For example, the attribute transformation unit (30007) can transform the attributes of a point at a position based on the position value of a point included in a voxel. As described above, when the position of the center point of a voxel is set based on the positions of one or more points included in a voxel, the attribute transformation unit (30007) transforms the attributes of one or more points. When trisoop geometry encoding is performed, the attribute conversion unit (30007) can convert attributes based on the trisoop geometry encoding.

[0090] The attribute transformation unit (30007) can perform attribute transformation by calculating the average value of attributes or attribute values ​​(e.g., the color or reflectance of each point) of neighboring points within a specific location / radius from the position (or position value) of the center point of each voxel. The attribute transformation unit (30007) can apply a weight based on the distance from the center point to each point when calculating the average value. Thus, each voxel has a position and a calculated attribute (or attribute value).

[0091] The attribute conversion unit (30007) can search for neighboring points within a specific location / radius from the position of the center point of each voxel based on a KD tree or a Molton code. A KD tree is a binary search tree that supports a data structure capable of managing points based on their positions to enable rapid Nearest Neighbor Search (NNS). A Molton code is generated by representing the coordinate values ​​(e.g., (x, y, z)) representing the 3D positions of all points as bit values ​​and mixing the bits. For example, if the coordinate values ​​representing the position of a point are (5, 9, 1), the bit values ​​of the coordinate values ​​are (0101, 1001, 0001). When the bit values ​​are mixed according to the bit indices in the order of z, y, and x, it becomes 010001000111. When this value is represented in decimal, it becomes 1095. That is, the Molton code value of the point with coordinates (5, 9, 1) is 1095. The attribute transformation unit (30007) sorts the points based on the Molton code value and can perform shortest neighbor search (NNS) through a depth-first traversal process. After the attribute transformation operation, if shortest neighbor search (NNS) is required in other transformation processes for attribute coding, a KD tree or Molton code is utilized.

[0092] As shown in the drawing, the converted attributes are input to the RAHT conversion unit (30008) and / or LOD generation unit (30009).

[0093] The RAHT transformation unit (30008) according to the embodiments performs RAHT coding to predict attribute information based on reconstructed geometry information. For example, the RAHT transformation unit (30008) can predict attribute information of a node at an upper level of the octree based on attribute information associated with a node at a lower level of the octree.

[0094] The LOD generation unit (30009) according to the embodiments generates a Level of Detail (LOD) to perform predictive transformation coding. The LOD according to the embodiments represents the degree of detail of the point cloud content, and indicates that the smaller the LOD value, the lower the detail of the point cloud content, and the larger the LOD value, the higher the detail of the point cloud content. Points can be classified according to the LOD.

[0095] The lifting transformation unit (30010) according to the embodiments performs lifting transformation coding that transforms the attributes of the point cloud based on weights. As described above, the lifting transformation coding may be applied optionally.

[0096] The coefficient quantization unit (30011) according to the embodiments quantizes attribute-coded attributes based on coefficients.

[0097] An arismetic encoder (30012) according to the embodiments encodes quantized attributes based on arismetic coding.

[0098] The elements of the point cloud encoder of FIG. 3 may be implemented in hardware, software, firmware, or a combination thereof, comprising one or more processors or integrated circuits configured to communicate with one or more memories included in the point cloud providing device, although not illustrated in the drawing. One or more processors may perform at least one of the operations and / or functions of the elements of the point cloud encoder of FIG. 3 described above. Additionally, one or more processors may operate or execute a set of software programs and / or instructions for performing the operations and / or functions of the elements of the point cloud encoder of FIG. 3. One or more memories according to the embodiments may include high-speed random access memory and may include non-volatile memory (e.g., one or more magnetic disk storage devices, flash memory devices, or other non-volatile solid-state memory devices).

[0099] FIG. 4 shows examples of octree and occupancy codes according to embodiments.

[0100] As described in FIGS. 1 to 3, a point cloud content providing system (point cloud video encoder (10002)) or a point cloud encoder (e.g., an octree analysis unit (30002)) performs octree geometry coding based on an octree structure (or octree coding) to efficiently manage the area and / or position of a voxel.

[0101] The top of FIG. 4 shows an octree structure. The three-dimensional space of the point cloud content according to the embodiments is represented by the axes of the coordinate system (e.g., X-axis, Y-axis, Z-axis). The octree structure has two poles (0,0,0) and (2 d , 2 d , 2 d It is generated by recursively subdividing the bounding box (cubical axis-aligned bounding box) defined by ). 2d can be set to the value that constitutes the smallest bounding box enclosing all points of the point cloud content (or point cloud video). d represents the depth of the octree. The value of d is determined according to the following equation. In the equation below, (x int n , y int n , z int n ) represents the positions (or position values) of quantized points.

[0102] d =Ceil(Log2(Max(x_n^int,y_n^int,z_n^int,n=1,…,N)+1))

[0103] As illustrated in the middle of the top of Fig. 4, the entire three-dimensional space can be divided into eight spaces according to the division. Each divided space is represented as a cube having six faces. As illustrated in the right of the top of Fig. 4, each of the eight spaces is further divided based on the axes of the coordinate system (e.g., X-axis, Y-axis, Z-axis). Thus, each space is again divided into eight smaller spaces. The divided smaller spaces are also represented as cubes having six faces. This division method is applied until the leaf nodes of the octree become voxels.

[0104] The bottom of Fig. 4 shows the occupancy code of an octree. The occupancy code of an octree is generated to indicate whether each of the eight partitioned spaces resulting from the partitioning of a single space contains at least one point. Therefore, one occupancy code is represented by eight child nodes. Each child node represents the occupancy of the partitioned space, and the child node has a value of 1 bit. Thus, the occupancy code is represented as an 8-bit code. That is, if the space corresponding to the child node contains at least one point, the node has a value of 1. If the space corresponding to the child node does not contain a point (empty), the node has a value of 0. Since the occupancy code shown in Fig. 4 is 00100001, it indicates that the spaces corresponding to the 3rd and 8th child nodes among the eight child nodes each contain at least one point. As illustrated in the drawing, the 3rd child node and the 8th child node each have 8 child nodes, and each child node is represented by an 8-bit Occupancy code. The drawing indicates that the Occupancy code of the 3rd child node is 10000111 and the Occupancy code of the 8th child node is 01001111. A point cloud encoder according to the embodiments (e.g., an arismetic encoder (30004)) can entropy-encode the Occupancy code. Additionally, to increase compression efficiency, the point cloud encoder can intra- / inter-encode the Occupancy code. A receiving device according to the embodiments (e.g., a receiving device (10004) or a point cloud video decoder (10006)) reconstructs the octree based on the Occupancy code.

[0105] A point cloud encoder according to the embodiments (e.g., the point cloud encoder of FIG. 3, or the octree analysis unit (30002)) can perform voxelization and octree coding to store the positions of the points. However, since points in a three-dimensional space are not always evenly distributed, there may be specific areas where few points exist. Therefore, performing voxelization on the entire three-dimensional space is inefficient. For example, if there are almost no points in a specific area, there is no need to perform voxelization up to that area.

[0106] Accordingly, the point cloud encoder according to the embodiments can perform direct coding, which directly codes the positions of points included in a specific region (or nodes excluding leaf nodes of an octree) without performing voxelization on the aforementioned specific region. The coordinates of the points directly coded according to the embodiments are referred to as the Direct Coding Mode (DCM). Additionally, the point cloud encoder according to the embodiments can perform trisoup geometry encoding, which reconstructs the positions of points within a specific region (or node) based on voxels using a surface model. Trisoup geometry encoding is a geometry encoding that represents an object as a series of triangle meshes. Therefore, the point cloud decoder can generate a point cloud from the mesh surface. Direct coding and trisoup geometry encoding according to the embodiments may be performed optionally. In addition, direct coding and trisoop geometry encoding according to the embodiments can be performed in combination with octree geometry coding (or octree coding).

[0107] To perform direct coding, the option to use direct mode for applying direct coding must be enabled, the node to which direct coding is to be applied must not be a leaf node, and there must be points within a specific node that are below a threshold. In addition, the total number of points subject to direct coding must not exceed a preset threshold. If the above conditions are satisfied, the point cloud encoder (or arismetic encoder (30004)) according to the embodiments can entropy-code the positions (or position values) of the points.

[0108] A point cloud encoder according to the embodiments (e.g., a surface approximation analysis unit (30003)) can determine a specific level of an octree (where the level is smaller than the depth d of the octree) and, starting from that level, perform trisoop geometry encoding to reconstruct the position of points within a node region based on voxels using a surface model (trisoop mode). The point cloud encoder according to the embodiments can specify the level to which trisoop geometry encoding is applied. For example, if the specified level is equal to the depth of the octree, the point cloud encoder does not operate in trisoop mode. That is, the point cloud encoder according to the embodiments can operate in trisoop mode only when the specified level is smaller than the depth value of the octree. A three-dimensional cubic region of nodes at a specified level according to the embodiments is referred to as a block. A block may include one or more voxels. A block or a voxel may correspond to a brick. Within each block, geometry is represented as a surface. A surface according to the embodiments may intersect each edge of the block at most once.

[0109] Since one block has 12 edges, there are at least 12 intersection points within one block. Each intersection point is referred to as a vertex. A vertex along an edge is detected if there is at least one occupied voxel adjacent to that edge among all blocks sharing that edge. An occupied voxel according to the embodiments means a voxel containing a point. The position of a vertex detected along an edge is the average position along the edge of all voxels adjacent to that edge among all blocks sharing that edge.

[0110] When a vertex is detected, the point cloud encoder according to the embodiments can entropy-code the edge start point (x, y, z), edge direction vector (Δx, Δy, Δz), and vertex position value (relative position value within the edge). When trisoop geometry encoding is applied, the point cloud encoder according to the embodiments (e.g., geometry reconstruction unit (30005)) can generate restored geometry (reconstructed geometry) by performing triangle reconstruction, up-sampling, and voxelization processes.

[0111] The vertices located on the edges of the block determine the surface passing through the block. The surface according to the embodiments is a non-planar polygon. The triangle reconstruction process reconstructs the surface represented by triangles based on the edge start point, the edge direction vector, and the vertex position value. The triangle reconstruction process is as follows: ① calculate the centroid value of each vertex, ② subtract the centroid value from each vertex value, ③ square the result, and add all the result together.

[0112]

[0113] The minimum sum is calculated, and a projection process is performed along the axis where the minimum value is located. For example, if the x-element is at its minimum, each vertex is projected along the x-axis relative to the center of the block and onto the (y, z) plane. If the value obtained from projecting onto the (y, z) plane is (ai, bi), the θ value is calculated using atan2(bi, ai), and the vertices are aligned based on the θ value. The table below shows the combinations of vertices to generate triangles depending on the number of vertices. The vertices are aligned in order from 1 to n. The table below indicates that for four vertices, two triangles can be formed depending on the combination of vertices. The first triangle can be formed from the 1st, 2nd, and 3rd vertices among the aligned vertices, and the second triangle can be formed from the 3rd, 4th, and 1st vertices among the aligned vertices.

[0114] Table 2-1. Triangles formed from vertices ordered 1,… ,n

[0115] n triangles

[0116] 3 (1,2,3)

[0117] 4 (1,2,3), (3,4,1)

[0118] 5 (1,2,3), (3,4,5), (5,1,3)

[0119] 6 (1,2,3), (3,4,5), (5,6,1), (1,3,5)

[0120] 7 (1,2,3), (3,4,5), (5,6,7), (7,1,3), (3,5,7)

[0121] 8 (1,2,3), (3,4,5), (5,6,7), (7,8,1), (1,3,5), (5,7,1)

[0122] 9 (1,2,3), (3,4,5), (5,6,7), (7,8,9), (9,1,3), (3,5,7), (7,9,3)

[0123] 10 (1,2,3), (3,4,5), (5,6,7), (7,8,9), (9,10,1), (1,3,5), (5,7,9), (9,1,5)

[0124] 11 (1,2,3), (3,4,5), (5,6,7), (7,8,9), (9,10,11), (11,1,3), (3,5,7), (7,9,11), (11,3,7)

[0125] 12 (1,2,3), (3,4,5), (5,6,7), (7,8,9), (9,10,11), (11,12,1), (1,3,5), (5,7,9), (9,11,1), (1,5,9)

[0126] The upsampling process is performed to voxelize by adding intermediate points along the edges of the triangle. Additional points are generated based on the upsampling factor value and the width of the block. The additional points are referred to as refined vertices. A point cloud encoder according to the embodiments can voxelize the refined vertices. Additionally, the point cloud encoder can perform attribute encoding based on the voxelized positions (or position values).

[0127] Figure 5 shows an example of a point configuration by LOD according to embodiments.

[0128] As described in FIGS. 1 to 4, the encoded geometry is reconstructed (decompressed) before attribute encoding is performed. When direct coding is applied, the geometry reconstruction operation may include changing the arrangement of the direct-coded points (e.g., placing the direct-coded points at the front of the point cloud data). When trisoop geometry encoding is applied, the geometry reconstruction process involves triangle reconstruction, upsampling, and voxelization. Since attributes depend on geometry, attribute encoding is performed based on the reconstructed geometry.

[0129] A point cloud encoder (e.g., an LOD generation unit (30009)) can reorganize points by LOD. The drawing shows point cloud content corresponding to the LOD. The left side of the drawing shows the original point cloud content. The second figure from the left of the drawing shows the distribution of points of the lowest LOD, and the rightmost figure of the drawing shows the distribution of points of the highest LOD. That is, the points of the lowest LOD are sparsely distributed, while the points of the highest LOD are densely distributed. In other words, according to the direction of the arrow indicated at the bottom of the drawing, as the LOD increases, the spacing (or distance) between points becomes shorter.

[0130] Figure 6 shows an example of a point configuration by LOD according to embodiments.

[0131] As described in FIGS. 1 to 5, a point cloud content providing system or a point cloud encoder (e.g., a point cloud video encoder (10002), the point cloud encoder of FIG. 3, or an LOD generation unit (30009)) can generate an LOD. The LOD is generated by reorganizing points into a set of refinement levels according to a set LOD distance value (or a set of Euclidean distances). The LOD generation process is performed in a point cloud decoder as well as a point cloud encoder.

[0132] The top of Fig. 6 shows examples of points (P0 to P9) of point cloud content distributed in three-dimensional space. The Original Order in Fig. 6 represents the order of points P0 to P9 prior to LOD generation. The LOD-based Order in Fig. 6 represents the order of points following LOD generation. Points are rearranged by LOD. Additionally, higher LODs include points belonging to lower LODs. As illustrated in Fig. 6, LOD0 includes P0, P5, P4, and P2. LOD1 includes the points of LOD0 and P1, P6, and P3. LOD2 includes the points of LOD0, the points of LOD1, and P9, P8, and P7.

[0133] As described in FIG. 3, the point cloud encoder according to the embodiments can perform predictive transform coding, lifting transform coding, and RAHT transform coding selectively or in combination.

[0134] The point cloud encoder according to the embodiments can generate predictors for points and perform predictive transformation coding to set the predicted attribute (or predicted attribute value) of each point. That is, N predictors can be generated for N points. The predictor according to the embodiments can calculate a weight (=1 / distance) value based on the LOD value of each point, indexing information for neighboring points within a set distance per LOD, and the distance value to the neighboring points.

[0135] According to the embodiments, the predicted attribute (or attribute value) is set as the average value of the values ​​obtained by multiplying the attributes (or attribute values, e.g., color, reflectance, etc.) of neighboring points set in the predictor of each point by a weight (or weight value) calculated based on the distance to each neighboring point. The point cloud encoder according to the embodiments (e.g., coefficient quantization unit (30011)) can quantize and inverse quantize the residual values ​​(which may be referred to as residual attributes, residual attribute values, attribute prediction residual values, etc.) obtained by subtracting the predicted attribute (attribute value) from the attribute (attribute value) of each point. The quantization process is as shown in the following table.

[0136] graph. Attribute prediction residuals quantization pseudo code

[0137] int PCCQuantization(int value, int quantStep) {

[0138] if( value >=0) {

[0139] return floor(value / quantStep + 1.0 / 3.0);

[0140] } else {

[0141] return -floor(-value / quantStep + 1.0 / 3.0);

[0142] }

[0143] }

[0144] graph. Attribute prediction residuals inverse quantization pseudo code

[0145] int PCCInverseQuantization(int value, int quantStep) {

[0146] if( quantStep ==0) {

[0147] return value;

[0148] } else {

[0149] return value * quantStep;

[0150] }

[0151] }

[0152] A point cloud encoder according to the embodiments (e.g., an arismetic encoder (30012)) can entropy-code the quantized and inversely quantized residual values ​​as described above when there are neighboring points in the predictor of each point. A point cloud encoder according to the embodiments (e.g., an arismetic encoder (30012)) can entropy-code the attributes of the corresponding point without performing the process described above when there are no neighboring points in the predictor of each point.

[0153] A point cloud encoder according to the embodiments (e.g., a lifting transformation unit (30010)) can perform lifting transformation coding by generating a predictor for each point, setting the LOD calculated in the predictor, registering neighboring points, and setting weights based on the distance to neighboring points. The lifting transformation coding according to the embodiments is similar to the prediction transformation coding described above, but differs in that weights are cumulatively applied to attribute values. The process of cumulatively applying weights to attribute values ​​according to the embodiments is as follows.

[0154] 1) Create an array QW (QuantizationWight) to store the weight values ​​of each point. The initial value of all elements in QW is 1.0. Add the value obtained by multiplying the current point's predictor weight by the QW value of the predictor index of the neighboring node registered in the predictor.

[0155] 2) Lift prediction process: To calculate the predicted attribute value, the value obtained by multiplying the point's attribute value by a weight is subtracted from the existing attribute value.

[0156] 3) Create temporary arrays named updateweight and update, and initialize the temporary arrays to 0.

[0157] 4) For all predictors, the calculated weight is additionally multiplied by the weight stored in the QW corresponding to the predictor index, and the resulting weight is accumulated in the update weight array with the neighbor node index. In the update array, the value obtained by multiplying the attribute value of the neighbor node index by the calculated weight is accumulated.

[0158] 5) Lift update process: For all predictors, the attribute value of the update array is divided by the weight value of the update weight array at the predictor index, and the original attribute value is added back to the divided value.

[0159] 6) For all predictors, the predicted attribute value is calculated by additionally multiplying the attribute value updated through the lift update process by the weight (stored in QW) updated through the lift prediction process. A point cloud encoder according to the embodiments (e.g., coefficient quantizer (30011)) quantizes the predicted attribute value. Additionally, a point cloud encoder (e.g., arismetic encoder (30012)) entropies the quantized attribute value.

[0160] A point cloud encoder according to the embodiments (e.g., a RAHT transform unit (30008)) can perform RAHT transform coding to predict attributes of upper-level nodes using attributes associated with nodes at lower levels of the octree. RAHT transform coding is an example of attribute intra-coding through octree backward scanning. A point cloud encoder according to the embodiments scans from voxels to the entire region and repeats the merging process up to the root node, merging voxels into larger blocks at each step. The merging process according to the embodiments is performed only on occupied nodes. The merging process is not performed on empty nodes, and the merging process is performed on the node immediately above the empty node.

[0161] The following equation represents the RAHT transformation matrix. g l x, y, z represents the average attribute value of the voxels at level l. g l x, y, z can be calculated from gl+1 2x, y, z and gl+1 2x+1, y, z. The weights of gl 2x, y, z and gl 2x+1, y, z are w1=wl 2x, y, z and w2=wl 2x+1, y, z.

[0162]

[0163] gl-1 x, y, z are low-pass values ​​used in the merging process at the next higher level. hl-1 x, y, z are high-pass coefficients, and the high-pass coefficients at each step are quantized and entropy-coded (e.g., encoding of an arismetic encoder (400012)). The weights are calculated as wl-1 x, y, z = wl 2x, y, z + wl 2x + 1, y, z. The root node is the last g 1 0, 0, 0 and g 1 0, 0, 1 It is generated as follows through.

[0164]

[0165] The gDC value is also quantized and entropy-coded, just like the high-pass coefficient.

[0166] FIG. 7 shows an example of a point cloud decoder according to embodiments.

[0167] The point cloud decoder illustrated in FIG. 7 is an example of a point cloud decoder and can perform a decoding operation, which is the reverse process of the encoding operation of the point cloud encoder described in FIG. 1 to 6.

[0168] As described in Fig. 1, the point cloud decoder can perform geometry decoding and attribute decoding. Geometry decoding is performed before attribute decoding.

[0169] A point cloud decoder according to the embodiments comprises an arithmetic decoder (7000), a synthesize octree (7001), a synthesize surface approximation (7002), a reconstruct geometry (7003), an inverse transform coordinates (7004), an arithmetic decoder (7005), an inverse quantize (7006), a RAHT transform (7007), an LOD generater (7008), an inverse lifting (7009), and / or an inverse transform colors (7010).

[0170] An arismetic decoder (7000), an octree composite unit (7001), a surface offset composite unit (7002), a geometry reconstruction unit (7003), and a coordinate system inverse transformation unit (7004) can perform geometry decoding. Geometry decoding according to the embodiments may include direct coding and trisoup geometry decoding. Direct coding and trisoup geometry decoding are applied optionally. Additionally, geometry decoding is not limited to the above examples and is performed as the reverse process of geometry encoding described in FIGS. 1 through 6.

[0171] The arismetic decoder (7000) according to the embodiments decodes the received geometry bitstream based on arismetic coding. The operation of the arismetic decoder (7000) corresponds to the reverse process of the arismetic encoder (30004).

[0172] The octree synthesis unit (7001) according to the embodiments can generate an octree by obtaining an Occupancy code from a decoded geometry bitstream (or information regarding the geometry obtained as a result of decoding). A specific description of the Occupancy code is as described in FIGS. 1 to 6.

[0173] The surface off-relation synthesis unit (7002) according to the embodiments can synthesize a surface based on the decoded geometry and / or the generated octree when trisoop geometry encoding is applied.

[0174] The geometry reconstruction unit (7003) according to the embodiments can regenerate geometry based on a surface and / or decoded geometry. As described in FIGS. 1 through 6, direct coding and trisoop geometry encoding are applied optionally. Accordingly, the geometry reconstruction unit (7003) directly retrieves and adds position information of points to which direct coding has been applied. In addition, when trisoop geometry encoding is applied, the geometry reconstruction unit (7003) can restore geometry by performing reconstruction operations of the geometry reconstruction unit (30005), such as triangle reconstruction, up-sampling, and voxelization operations. Specific details are omitted as they are the same as those described in FIG. 4. The restored geometry may include a point cloud picture or frame that does not contain attributes.

[0175] The coordinate system inverse transformation unit (7004) according to the embodiments can obtain the positions of the points by transforming the coordinate system based on the restored geometry.

[0176] The arismetic decoder (7005), inverse quantization unit (7006), RAHT transformation unit (7007), LOD generation unit (7008), inverse lifting unit (7009), and / or color inverse transformation unit (7010) can perform attribute decoding as described in FIG. 10. Attribute decoding according to the embodiments may include Region Adaptive Hierarchial Transform (RAHT) decoding, Interpolaration-based hierarchical nearest-neighbour prediction-Prediction Transform) decoding, and interpolation-based hierarchical nearest-neighbour prediction with an update / lifting step (Lifting Transform) decoding. The three decodings described above may be used optionally, or a combination of one or more decodings may be used. Furthermore, attribute decoding according to the embodiments is not limited to the examples described above.

[0177] The arismetic decoder (7005) according to the embodiments decodes the attribute bitstream into arismetic coding.

[0178] The inverse quantization unit (7006) according to the embodiments inverse quantizes information about the decoded attribute bitstream or the attribute obtained as a result of decoding and outputs the inverse quantized attributes (or attribute values). Inverse quantization may be optionally applied based on the attribute encoding of the point cloud encoder.

[0179] According to embodiments, the RAHT transformation unit (7007), LOD generation unit (7008), and / or inverse lifting unit (7009) can process the reconstructed geometry and inverse quantized attributes. As described above, the RAHT transformation unit (7007), LOD generation unit (7008), and / or inverse lifting unit (7009) can optionally perform a corresponding decoding operation according to the encoding of the point cloud encoder.

[0180] The color inverse conversion unit (7010) according to the embodiments performs inverse conversion coding to inversely convert the color value (or texture) included in the decoded attributes. The operation of the color inverse conversion unit (7010) may be selectively performed based on the operation of the color conversion unit (30006) of the point cloud encoder.

[0181] The elements of the point cloud decoder of FIG. 7 may be implemented in hardware, software, firmware, or a combination thereof, comprising one or more processors or integrated circuits configured to communicate with one or more memories included in the point cloud providing device, although not illustrated in the drawing. One or more processors may perform at least one of the operations and / or functions of the elements of the point cloud decoder of FIG. 7 described above. Additionally, one or more processors may operate or execute a set of software programs and / or instructions for performing the operations and / or functions of the elements of the point cloud decoder of FIG. 7.

[0182] FIG. 8 is an example of a transmission device according to embodiments.

[0183] The transmission device illustrated in FIG. 8 is an example of the transmission device (10000) of FIG. 1 (or the point cloud encoder of FIG. 3). The transmission device illustrated in FIG. 8 can perform at least one of the same or similar operations and methods as the operations and encoding methods of the point cloud encoder described in FIG. 1 to 6. A transmission device according to embodiments may include a data input unit (8000), a quantization processing unit (8001), a voxelization processing unit (8002), an octree occupancy code generation unit (8003), a surface model processing unit (8004), an intra / inter coding processing unit (8005), an arithmetic coder (8006), a metadata processing unit (8007), a color conversion processing unit (8008), an attribute conversion processing unit (or attribute conversion processing unit) (8009), a prediction / lifting / RAHT conversion processing unit (8010), an arithmetic coder (8011) and / or a transmission processing unit (8012).

[0184] The data input unit (8000) according to the embodiments receives or acquires point cloud data. The data input unit (8000) may perform an operation and / or acquisition method identical or similar to the operation and / or acquisition method of the point cloud video acquisition unit (10001) (or the acquisition process (20000) described in FIG. 2).

[0185] The data input unit (8000), quantization processing unit (8001), voxelization processing unit (8002), octree occupancy code generation unit (8003), surface model processing unit (8004), intra / inter coding processing unit (8005), and arithmetic coder (8006) perform geometry encoding. Since the geometry encoding according to the embodiments is identical or similar to the geometry encoding described in FIGS. 1 to 6, a detailed description is omitted.

[0186] The quantization processing unit (8001) according to the embodiments quantizes geometry (e.g., location values ​​of points, or position values). The operation and / or quantization of the quantization processing unit (8001) is the same or similar to the operation and / or quantization of the quantization unit (30001) described in FIG. 3. The specific description is the same as that described in FIG. 1 through 6.

[0187] The voxelization processing unit (8002) according to the embodiments voxelizes the position values ​​of the quantized points. The voxelization processing unit (80002) may perform the same or similar operation and / or process as the operation and / or voxelization process of the quantization unit (30001) described in FIG. 3. The specific description is the same as that described in FIG. 1 to 6.

[0188] The octree occupancy code generation unit (8003) according to the embodiments performs octree coding on the positions of voxelized points based on an octree structure. The octree occupancy code generation unit (8003) can generate an occupancy code. The octree occupancy code generation unit (8003) can perform operations and / or methods identical or similar to the operations and / or methods of the point cloud encoder (or octree analysis unit (30002)) described in FIGS. 3 and 4. The specific description is the same as that described in FIGS. 1 through 6.

[0189] The surface model processing unit (8004) according to the embodiments can perform trisup geometry encoding that reconstructs the positions of points within a specific region (or node) based on a voxel based on a surface model. The surface model processing unit (8004) can perform operations and / or methods identical or similar to the operations and / or methods of the point cloud encoder (e.g., surface approximation analysis unit (30003)) described in FIG. 3. The specific description is the same as that described in FIG. 1 through 6.

[0190] According to the embodiments, the intra / inter coding processing unit (8005) can intra / inter code point cloud data. The intra / inter coding processing unit (8005) can perform coding identical or similar to the intra / inter coding described in FIG. 7. The specific description is the same as that described in FIG. 7. According to the embodiments, the intra / inter coding processing unit (8005) may be included in an arismetic coder (8006).

[0191] An arismetic coder (8006) according to the embodiments entropy-encodes an octree and / or approximated octree of point cloud data. For example, the encoding method includes an arismetic encoding method. The arismetic coder (8006) performs the same or similar operation and / or method as the arismetic encoder (30004).

[0192] A metadata processing unit (8007) according to the embodiments processes metadata regarding point cloud data, such as setting values, and provides it to necessary processing processes such as geometry encoding and / or attribute encoding. Additionally, a metadata processing unit (8007) according to the embodiments may generate and / or process signaling information related to geometry encoding and / or attribute encoding. The signaling information according to the embodiments may be encoded separately from geometry encoding and / or attribute encoding. Additionally, the signaling information according to the embodiments may be interleaved.

[0193] The color conversion processing unit (8008), attribute conversion processing unit (8009), prediction / lifting / RAHT conversion processing unit (8010), and arithmetic coder (8011) perform attribute encoding. Since the attribute encoding according to the embodiments is identical or similar to the attribute encoding described in FIGS. 1 to 6, a detailed description is omitted.

[0194] The color conversion processing unit (8008) according to the embodiments performs color conversion coding that converts color values ​​included in attributes. The color conversion processing unit (8008) may perform color conversion coding based on reconstructed geometry. The description of the reconstructed geometry is the same as that described in FIGS. 1 through 6. In addition, it performs the same or similar operation and / or method as the operation and / or method of the color conversion unit (30006) described in FIG. 3. A detailed description is omitted.

[0195] The attribute transformation processing unit (8009) according to the embodiments performs attribute transformation that transforms attributes based on positions where geometry encoding has not been performed and / or reconstructed geometry. The attribute transformation processing unit (8009) performs operations and / or methods identical or similar to the operations and / or methods of the attribute transformation unit (30007) described in FIG. 3. A detailed description is omitted. The prediction / lifting / RAHT transformation processing unit (8010) according to the embodiments may code the transformed attributes by RAHT coding, prediction transformation coding, and lifting transformation coding, or a combination thereof. The prediction / lifting / RAHT transformation processing unit (8010) performs at least one of operations identical or similar to the operations of the RAHT transformation unit (30008), LOD generation unit (30009), and lifting transformation unit (30010) described in FIG. 3. In addition, the descriptions of predictive transformation coding, lifting transformation coding, and RAHT transformation coding are the same as those described in Figures 1 to 6, so a detailed description is omitted.

[0196] The arismetic coder (8011) according to the embodiments can encode coded attributes based on arismetic coding. The arismetic coder (8011) performs the same or similar operation and / or method as the operation and / or method of the arismetic encoder (300012).

[0197] A transmission processing unit (8012) according to embodiments may transmit each bitstream containing encoded geometry and / or encoded attributes and metadata information, or may transmit the encoded geometry and / or encoded attributes and metadata information by configuring them into a single bitstream. When the encoded geometry and / or encoded attributes and metadata information according to embodiments is configured into a single bitstream, the bitstream may include one or more sub-bitstreams. The bitstream according to embodiments may include signaling information and slice data, including SPS (Sequence Parameter Set) for sequence-level signaling, GPS (Geometry Parameter Set) for signaling of geometry information coding, APS (Attribute Parameter Set) for signaling of attribute information coding, and TPS (Tile Parameter Set) for tile-level signaling. The slice data may include information about one or more slices. One slice according to embodiments is one geometry bitstream (Geom0 0 ) and one or more attribute bitstreams (Attr0 0 , Attr1 0 It may include ).

[0198] A slice refers to a series of syntax elements representing all or part of a coded point cloud frame.

[0199] According to the embodiments, the TPS may include information regarding each tile (e.g., coordinate value information of a bounding box and height / size information, etc.) for one or more tiles. The geometry bitstream may include a header and a payload. The header of the geometry bitstream according to the embodiments may include identification information of a parameter set included in the GPS (geom_parameter_set_id), a tile identifier (geom_tile_id), a slice identifier (geom_slice_id), and information regarding data included in the payload, etc. As described above, the metadata processing unit (8007) according to the embodiments may generate and / or process signaling information and transmit it to the transmission processing unit (8012). According to the embodiments, the elements performing geometry encoding and the elements performing attribute encoding may share data / information with each other as indicated by the dotted lines. The transmission processing unit (8012) according to the embodiments may perform an operation and / or transmission method identical or similar to the operation and / or transmission method of the transmitter (10003). A detailed explanation is omitted as it is the same as that described in FIGS. 1 and 2.

[0200] FIG. 9 is an example of a receiving device according to embodiments.

[0201] The receiving device illustrated in FIG. 9 is an example of the receiving device (10004) of FIG. 1 (or the point cloud decoder of FIG. 10 and FIG. 11). The receiving device illustrated in FIG. 9 can perform at least one of the same or similar operations and methods as the operations and decoding methods of the point cloud decoder described in FIG. 1 to FIG. 11.

[0202] A receiving device according to the embodiments may include a receiving unit (9000), a receiving processing unit (9001), an arithmetic decoder (9002), an occupancy code-based octree reconstruction processing unit (9003), a surface model processing unit (triangle reconstruction, up-sampling, voxelization) (9004), an inverse quantization processing unit (9005), a metadata parser (9006), an arithmetic decoder (9007), an inverse quantization processing unit (9008), a prediction / lifting / RAHT inverse transformation processing unit (9009), a color inverse transformation processing unit (9010), and / or a renderer (9011). Each component of the decoding according to the embodiments may perform the inverse process of the components of the encoding according to the embodiments.

[0203] A receiver (9000) according to the embodiments receives point cloud data. The receiver (9000) may perform an operation and / or a receiving method identical or similar to the operation and / or receiving method of the receiver (10005) of FIG. 1. A detailed description is omitted.

[0204] A receiving processing unit (9001) according to the embodiments can obtain a geometry bitstream and / or an attribute bitstream from the received data. The receiving processing unit (9001) may be included in the receiving unit (9000).

[0205] The arismetic decoder (9002), the Occupancy code-based octree reconstruction processing unit (9003), the surface model processing unit (9004), and the inverse quantization processing unit (9005) can perform geometry decoding. Since the geometry decoding according to the embodiments is identical or similar to the geometry decoding described in FIGS. 1 to 10, a detailed description is omitted.

[0206] The arismetic decoder (9002) according to the embodiments can decode a geometry bitstream based on arismetic coding. The arismetic decoder (9002) performs the same or similar operation and / or coding as the operation and / or coding of the arismetic decoder (7000).

[0207] According to the embodiments, the Occupancy code-based octree reconstruction processing unit (9003) can reconstruct an octree by obtaining an Occupancy code from a decoded geometry bitstream (or information regarding geometry obtained as a result of decoding). The Occupancy code-based octree reconstruction processing unit (9003) performs the same or similar operations and / or methods as the octree synthesis unit (7001) and / or octree generation method. According to the embodiments, the surface model processing unit (9004) can perform trisup geometry decoding and related geometry reconstruction (e.g., triangle reconstruction, up-sampling, voxelization) based on the surface model method when trisup geometry encoding is applied. The surface model processing unit (9004) performs the same or similar operations as the surface offset synthesis unit (7002) and / or geometry reconstruction unit (7003).

[0208] The inverse quantization processing unit (9005) according to the embodiments can inverse quantize the decoded geometry.

[0209] A metadata parser (9006) according to the embodiments can parse metadata included in the received point cloud data, such as setting values, etc. The metadata parser (9006) can pass the metadata to geometry decoding and / or attribute decoding. A specific description of the metadata is omitted as it is the same as described in FIG. 8.

[0210] The arismetic decoder (9007), inverse quantization processing unit (9008), prediction / lifting / RAHT inverse transformation processing unit (9009), and color inverse transformation processing unit (9010) perform attribute decoding. Since attribute decoding is identical or similar to the attribute decoding described in FIGS. 1 to 10, a detailed description is omitted.

[0211] The arismetic decoder (9007) according to the embodiments can decode an attribute bitstream into arismetic coding. The arismetic decoder (9007) can perform decoding of the attribute bitstream based on reconstructed geometry. The arismetic decoder (9007) performs the same or similar operation and / or coding as the operation and / or coding of the arismetic decoder (7005).

[0212] The inverse quantization processing unit (9008) according to the embodiments can inverse quantize the decoded attribute bitstream. The inverse quantization processing unit (9008) performs the same or similar operation and / or method as the operation and / or inverse quantization method of the inverse quantization unit (7006).

[0213] According to the embodiments, the prediction / lifting / RAHT inverse transformation processing unit (9009) can process the reconstructed geometry and inverse quantized attributes. The prediction / lifting / RAHT inverse transformation processing unit (9009) performs at least one of the same or similar operations and / or decodings as the operations and / or decodings of the RAHT transformation unit (7007), LOD generation unit (7008), and / or inverse lifting unit (7009). According to the embodiments, the color inverse transformation processing unit (9010) performs inverse transformation coding to inversely transform the color values ​​(or textures) included in the decoded attributes. The color inverse transformation processing unit (9010) performs the same or similar operations and / or inverse transformation coding as the operations and / or inverse transformation coding of the color inverse transformation unit (7010). A renderer (9011) according to the embodiments can render point cloud data.

[0214] FIG. 10 shows an example of a structure that can be linked with a point cloud data transmission / reception method / device according to embodiments.

[0215] The structure of FIG. 10 represents a configuration in which at least one of a server (1060), a robot (1010), an autonomous vehicle (1020), an XR device (1030), a smartphone (1040), a home appliance (1050) and / or an HMD (1070) is connected to a cloud network (1010). The robot (1010), the autonomous vehicle (1020), the XR device (1030), the smartphone (1040), or the home appliance (1050) are referred to as devices. Additionally, the XR device (1030) may correspond to a point cloud data (PCC) device according to the embodiments or may be linked with a PCC device.

[0216] The cloud network (1000) may refer to a network that constitutes part of the cloud computing infrastructure or exists within the cloud computing infrastructure. Here, the cloud network (1000) may be configured using a 3G network, a 4G or LTE (Long Term Evolution) network, or a 5G network, etc.

[0217] The server (1060) is connected to at least one of a robot (1010), an autonomous vehicle (1020), an XR device (1030), a smartphone (1040), a home appliance (1050) and / or an HMD (1070) via a cloud network (1000) and can assist in at least some of the processing of the connected devices (1010 to 1070).

[0218] The HMD (Head-Mount Display) (1070) represents one of the types in which an XR device and / or PCC device according to the embodiments may be implemented. A device of the HMD type according to the embodiments includes a communication unit, a control unit, a memory unit, an I / O unit, a sensor unit, and a power supply unit, etc.

[0219] Hereinafter, various embodiments of the device (1010 to 1050) to which the above-described technology is applied are described. Here, the device (1010 to 1050) illustrated in FIG. 10 may be linked / coupled with a point cloud data transmission / reception device according to the above-described embodiments.

[0220] <PCC+XR>

[0221] The XR / PCC device (1030) may be implemented as a Head-Mount Display (HMD), a Head-Up Display (HUD) equipped in a vehicle, a television, a mobile phone, a smartphone, a computer, a wearable device, a home appliance, digital signage, a vehicle, a stationary robot, or a mobile robot by applying PCC and / or XR (AR+VR) technology.

[0222] The XR / PCC device (1030) can obtain information about surrounding space or real objects by analyzing 3D point cloud data or image data obtained through various sensors or from an external device to generate position data and attribute data for 3D points, and can render and output an XR object to be output. For example, the XR / PCC device (1030) can output an XR object containing additional information about a recognized object by associating it with the recognized object.

[0223] <PCC+XR+모바일폰>

[0224] The XR / PCC device (1030) can be implemented as a mobile phone (1040) or the like by applying PCC technology.

[0225] The mobile phone (1040) can decode and display point cloud content based on PCC technology.

[0226] <PCC+자율주행+XR>

[0227] The autonomous vehicle (1020) can be implemented as a mobile robot, vehicle, unmanned aerial vehicle, etc. by applying PCC technology and XR technology.

[0228] An autonomous vehicle (1020) equipped with XR / PCC technology may refer to an autonomous vehicle equipped with means for providing XR images, or an autonomous vehicle that is the subject of control / interaction within the XR images. In particular, the autonomous vehicle (1020) that is the subject of control / interaction within the XR images is distinguished from the XR device (1030) and can be interconnected with it.

[0229] An autonomous vehicle (1020) equipped with means for providing XR / PCC images can acquire sensor information from sensors including cameras and output XR / PCC images generated based on the acquired sensor information. For example, the autonomous vehicle (1020) can provide an XR / PCC object corresponding to a real object or an object in the screen to the occupant by providing an XR / PCC object by outputting an XR / PCC image with a HUD.

[0230] At this time, when the XR / PCC object is displayed on the HUD, at least a portion of the XR / PCC object may be displayed so as to overlap with the actual object to which the occupant's gaze is directed. On the other hand, when the XR / PCC object is displayed on a display provided inside the autonomous vehicle, at least a portion of the XR / PCC object may be displayed so as to overlap with an object on the screen. For example, the autonomous vehicle (1220) may display XR / PCC objects corresponding to objects such as lanes, other vehicles, traffic lights, traffic signs, motorcycles, pedestrians, buildings, etc.

[0231] VR (Virtual Reality) technology, AR (Augmented Reality) technology, MR (Mixed Reality) technology and / or PCC (Point Cloud Compression) technology according to the embodiments can be applied to various devices.

[0232] In other words, VR technology is a display technology that provides real-world objects or backgrounds solely as CG images. On the other hand, AR technology refers to a technology that displays virtual CG images alongside images of real objects. Furthermore, MR technology is similar to the aforementioned AR technology in that it mixes and combines virtual objects with the real world. However, it is distinguished from AR technology in that while AR technology maintains a clear distinction between real-world objects and virtual objects created from CG images, using virtual objects to complement real-world objects, MR technology regards virtual objects as having the same nature as real-world objects. To give a more specific example, the aforementioned MR technology is applied in hologram services.

[0233] However, recently, rather than clearly distinguishing between VR, AR, and MR technologies, they are also referred to as XR (extended Reality) technology. Therefore, embodiments of the present invention are applicable to all VR, AR, MR, and XR technologies. These technologies may utilize encoding / decoding based on PCC, V-PCC, and G-PCC technologies.

[0234] The PCC method / device according to the embodiments can be applied to a vehicle providing autonomous driving services.

[0235] Vehicles providing autonomous driving services are connected to PCC devices to enable wired / wireless communication.

[0236] When a point cloud data (PCC) transceiver according to the embodiments is connected to a vehicle for wired or wireless communication, it can receive and process content data related to AR / VR / PCC services that can be provided along with an autonomous driving service, and transmit it to the vehicle. Additionally, when the point cloud data transceiver is mounted on a vehicle, the point cloud transceiver can receive and process content data related to AR / VR / PCC services according to a user input signal received through a user interface device and provide it to the user. A vehicle or a user interface device according to the embodiments can receive a user input signal. The user input signal according to the embodiments may include a signal indicating an autonomous driving service.

[0237] A point cloud data transmission method / device (or encoding method and device) according to embodiments may include a transmission device (10000) of FIG. 1, a point cloud video encoder (10002), a transmitter (10003), an acquisition-encoding-transmission (20000-20001-20002) of FIG. 2, encoding based on FIG. 11 to 29, generation of syntax information of FIG. 30 to 32, filtering of FIG. 33 to 35, generation of syntax information of FIG. 36a and 36ba and FIG. 36a and 36bb to 37, an encoder of FIG. 38, an encoding method of FIG. 40, etc.

[0238] A method / device for receiving point cloud data (or a decoding method and device) according to embodiments may include a receiving device (10004) of FIG. 1, a receiver (10005), a point cloud video decoder (10006), a transmission-decoding-rendering (20002-20003-20004) of FIG. 2, a decoder of FIG. 7, a receiving device of FIG. 9, a device of FIG. 10, encoding based on FIG. 11 to 29, acquisition of syntax information of FIG. 30 to 32, filtering of FIG. 33 to 35, acquisition of syntax information of FIG. 36a and FIG. 36b to 37, a decoder of FIG. 39, a decoding method of FIG. 41, etc.

[0239] In addition, the point cloud data transmission / reception method / device according to the embodiments may be referred to simply as the method / device according to the embodiments.

[0240] According to the embodiments, geometry data, geometry information, location information, etc. constituting the point cloud data are interpreted as having the same meaning. Attribute data, attribute information, attribute information, etc. constituting the point cloud data are interpreted as having the same meaning.

[0241] The method and apparatus according to the embodiments may include and perform an apparatus for encoding and decoding attributes of G-PCC files using fast nearest neighbor search and optimal window size determination.

[0242] The embodiments include methods for improving the attribute representation of G-PCC files. The described methods may be described based on ISO-based media file formats for G-PCC transmission.

[0243] The embodiments relate to a method for increasing the compression efficiency of Geometry-based Point Cloud Compression (G-PCC) for compressing three-dimensional point cloud data.

[0244] A point cloud consists of a set of points, and each point can have geometry information and attribute information. Geometry information is 3D position (XYZ) information, and attribute information is color (RGB, YUV, etc.) or / and reflection values.

[0245] The G-PCC encoding process can divide the point cloud into tiles according to region and divide each tile into slices for parallel processing. It can be composed of a process of compressing geometry on a slice-by-slice basis and compressing attribute information based on reconstructed geometry (decoded geometry) with location information changed through compression.

[0246] The G-PCC decoding process can be composed of receiving a geometry bitstream and an attribute bitstream in the form of encoded slice units, decoding the geometry, and decoding attribute information based on the geometry reconstructed through the decoding process.

[0247] Octree-based, predictive tree-based, or trisoup-based compression techniques can be used to compress geometry information.

[0248] For attribute information compression, predicting transform-based, lifting transform-based, or RAHT transform-based compression techniques can be used.

[0249] The embodiments relate to a technique for applying a filter, for example, an Adaptive Loop Filter (ALF), to increase the compression rate of point cloud attributes. The embodiments are described based on the ISO / IEC 23090-9 standard document.

[0250] A draft specification for transmitting G-PCC data in the ISO-based media file format (ISOBMF) has progressed to the Committee Draft (CD) status. Point clouds are widely used for several important reasons, including 3D representation, versatility, architecture and construction, geospatial and mapping, manufacturing and quality control, entertainment and media, medical imaging, precision and accuracy, data richness, interoperability, automation, and efficiency.

[0251] Searching for adjacent voxels in a point cloud is a fundamental task in various application fields. For example, it is as follows.

[0252] Surface reconstruction: Reconstructs a continuous surface from a set of discontinuous points.

[0253] Noise Reduction: Improves the quality of point cloud data by removing noise.

[0254] Partitioning and Clustering: Divides the point cloud into meaningful segments or clusters.

[0255] Compression Algorithm: Utilizes the spatial consistency of adjacent voxels to reduce redundancy and implement more efficient storage.

[0256] Filtering: Improves quality and enhances compression by using pre / post-processing, loop filters, motion correction with interpolation filters, or optical flow.

[0257] Modern video coding methods universally incorporate loop filters at the final stage of frame or slice encoding. For example, AVC / H.264 uses a deblocking filter[1], whereas HEVC / H.265 uses both a deblocking filter and a sample adaptive offset (SAO). HEVC's deblocking filter offers two variations, weak and strong, while SAO offers two types: edge offset and band offset. In VVC / H.266, both the adaptive loop filter (ALF) and SAO have been adopted. Furthermore, these concepts are extended to cross-component filtering, such as the cross-component adaptive loop filter (CC-ALF)[4] and cross-color filtering between layers for scalability[5]. On the encoder side, determining filter coefficients typically involves linear regression methods. Therefore, cross-component linear model prediction is often discussed alongside loop filters in this context, even though special filtering for neighboring pixels is not required. To improve attribute coding efficiency, G-PCC encoders / decoders can utilize adaptive loop filters. To achieve high levels of adaptability, flexibility, compression gain, and parallelism, adaptive loop filters are not applied only once per individual frame or slice, but rather as a set of multiple filters tailored to selected regions. The size of these regions can vary depending on conditions such as quantization parameters (QP), sparsity, and other factors. These filters can also function as post-filters. However, if the encoder can control the ALF parameters, it can guarantee the desired level of quality and pass this information to the decoder. Additionally, loop filters improve the quality of the reference frame, which helps increase the efficiency of inter-prediction and reduce bit size. While post-filtering requires an additional buffer, no additional memory is needed because the output signal after filtering can become part of the decoder picture buffer (DPB). An additional feature of region-based adaptive loop filters is the transmission of an "on / off" flag to the decoder.Therefore, the encoder can determine whether to apply ALF to all small areas based on Rate Distortion Optimization criteria or attention (salvage) maps. Finally, the loop filter is a well-known tool that improves visual quality by removing artifacts such as blocking and ringing. Therefore, this process must be prescriptive.

[0258] The definitions of terms according to the embodiments are as follows.

[0259] - Point Cloud Frame: A set of 3D points specified by Cartesian coordinates (x, y, z) and a fixed set of corresponding attributes at a specific point in time (optional)

[0260] - Bounding box: A rectangular prism containing the source point cloud frame

[0261] - Geometry: A set of Cartesian coordinates associated with a point cloud frame

[0262] - Properties: Scalar or vector properties associated with each point in the point cloud, such as color, reflectance, frame index (optional)

[0263] - APS: Set of attribute parameters

[0264] - ASH: Attribute Slice Header

[0265] - GSH: Geometry Slice Header

[0266] - GPS: Set of geometry parameters

[0267] - LSB: Lowest bit

[0268] - RAHT: Domain Adaptive Layer Transformation

[0269] - SPS: Sequence parameter set

[0270] - TPS: Set of tile parameters (same as tile inventory)

[0271] - Slice: A series of syntactic elements representing part or all of a coded point cloud frame

[0272] - 3D Tile: A rectangular prism inside a bounding box

[0273] - ALF: Adaptive Loop Filter

[0274] - RDO: Rate Distortion Optimization

[0275] - HLS: High-level syntax

[0276] - QP: Quantization parameter

[0277] Below, Example 1:

[0278] The fundamental concept of the Adaptive Loop Filter (ALF) is to minimize the mean squared error between the properties of the original voxels and the decoded voxels through convolution operations using Wiener-based adaptive filter coefficients. First developed by Norbert Wiener in 1940, this technique is applied at the final processing stage of each image to effectively remove visual artifacts from previous stages. The encoder determines the optimal filter coefficients and explicitly passes them to the decoder. Local adaptation techniques, which apply different filters to multiple regions within the image, are used to increase coding efficiency, particularly in low quantization rate (QP) scenarios. Furthermore, coding efficiency is further enhanced through the "on / off" control of the filter in each selected region. To solve this problem, the Wiener-Hopf equations must be solved in matrix form. This system of linear equations can be solved using Gaussian elimination (Gauss's algorithm). However, since matrices are symmetric, more efficient solutions can be obtained using special techniques such as Cholesky decomposition or conjugate gradient methods. Assuming symmetry of coding errors, it is wise to introduce symmetric filters. Such filters can reduce both the number of multiplications and the overhead of transmission coefficients by halving the number of filter coefficients. Considering that multipliers occupy significantly more chip area than adders, reducing the number of multipliers greatly reduces the chip area required for the ALF.

[0279] 1) In terms of the encoder, the application of a region-based adaptive loop filter consists of the following key steps.

[0280] Divide the point cloud into cubes of size 2p x 2p x 2p.

[0281] Search for neighbors for the processed voxels within each cube and generate a distribution for the corresponding tabs.

[0282] Calculate the matrix and dominance vector of the Wiener-Hopf equation using the original and reconstructed attribute signals.

[0283] Solve the Wiener-Hopf equation using floating-point operations.

[0284] Scaling factor and rounding.

[0285] Calculate convolution filter coefficients using the attribute values ​​of the referenced voxels.

[0286] The decision to apply is made based on RDO or user salience maps.

[0287] Write high-level syntax flags (whether ALF needs to be applied)

[0288] Write the overhead (integer ALF coefficients) using entropy coding.

[0289] 2) The application of the region-based adaptive loop filter on the decoder side consists of the following key steps.

[0290] Read HLS flags (whether ALF needs to be applied)

[0291] If the ALF flag is true, read the data required from the bitstream for ALF processing for each area (2p x 2p x 2p cube).

[0292] Divide the point cloud into 2p x 2p x 2p cubes.

[0293] If the current region is processed as ALF, search for the neighbors of the processed voxels within each cube.

[0294] Calculate convolution filter coefficients using the attribute values ​​of the referenced voxels.

[0295] For each processed voxel, neighbor voxel search and convolution calculations can be performed in parallel.

[0296] Below, Example 2:

[0297] FIG. 11 is an example of point cloud segmentation according to embodiments.

[0298] The method according to the embodiments includes the following steps.

[0299] FIG. 11(a) illustrates an example of point cloud segmentation for a 256x256x256 cube, and FIG. 11(b) illustrates an example of point cloud segmentation for a 512x512x512 cube and cube boundaries.

[0300] Step 1) Partition

[0301] The entire region containing a given point cloud can be divided into separate cubes of size 2p. When the cube size is a power of 2, determining whether a given voxel is contained in a specific cube becomes simple. This can be performed efficiently using bitwise shift operations.

[0302] If the intersection of the current cube and the given point cloud contains a sufficient number of voxels (i.e., if the number of voxels exceeds a predefined threshold), the encoder generates a temporary point cloud for both the reconstructed image and the original image. This temporary point cloud consists only of voxels belonging to the intersection.

[0303] If the voxel order of the original cloud differs from that of the reconstructed cloud, the order of the temporary cloud containing the original attributes must be reordered. However, in most cases, the original signal point cloud maintains the same order as the reconstructed point cloud. The size of the cube significantly affects the compression gain. Larger cubes reduce the overhead of the ALF data, while smaller sizes improve quality. Reducing overhead and improving quality are often conflicting goals. For high QPs, the number of bits is more critical, and overhead must be small. For low QPs, quality is more critical than the number of bits. Various experiments using different QP values ​​and sequences demonstrate the following optimal relationships.

[0304] QPpSIZE = 2 p 511120484611204840101024349512

[0305] Table 1 shows the dependence of cube size on QP. Therefore, for a size of 2p with base 2 and exponent p, this dependence can be approximated with high accuracy through the following equation: p=

[0306] For Table 1, it is as follows: p=(QP-34) / 6+9

[0307] FIG. 12 shows an example in which an original signal and a reconstructed signal according to embodiments are contained in a single cube.

[0308] Step 2) Search for Neighbors

[0309] Find the neighbors of the processed voxels within the 2-sided cube and find the distribution for the corresponding tabs.

[0310] The primary goal of this step is to identify neighboring voxels and utilize their attribute values ​​in the convolution calculation. This aims to minimize the difference between the reconstructed signal and the original signal using the Mean Squared Error (MSE) as the cost function: ||Org-Filter* Rec||->min

[0311] This process represents the most complex and sensitive aspect. Since a simple search for neighbors operates as a quadratic function of the number of voxels, it results in an extremely labor-intensive task on a cube of size 2048. Therefore, the method according to the embodiments utilizes Molten code information to increase the search speed. Another issue is the criteria for selecting neighbor sizes. Two images Egyptian_mask_vox12 and soldier_vox10_0690 are shown in FIG. 13 at the same scale.

[0312] FIG. 13 shows an example in which two different point clouds of the same size and two spheres with radii of 16 and 2 are added for comparison according to embodiments.

[0313] To balance quality improvement with overhead reduction, it is essential to use a relatively small number of taps and limit the number of adjacent voxels for each processed voxel. Therefore, the window size for determining the adjacent regions to locate voxels must be selected individually for each point cloud. The window size varies significantly depending on the geometric characteristics representing the voxel distribution within the cloud. For sparse clouds, the window size may be relatively large, whereas for high-density clouds, it must be smaller. Consequently, the window size is directly influenced by voxel density. While voxel density can be calculated on both the encoder and decoder sides for each cube, this parameter is generally stable for most point clouds and can be transmitted as a bitstream after being determined for an entire frame or slice.

[0314] FIG. 14 shows examples of Z-order curves (Lebesgue curve 1904, Morton space-filling curve 1966) according to embodiments.

[0315] Acceleration method for neighbor foxhole search based on Morton code information according to embodiments:

[0316] To form filter tabs, all voxels within the spherical neighborhood of the current voxel must be identified. Finding these neighboring voxels (and forming a general Wiener-Hopf system) is the bottleneck of filtering algorithms. The computational complexity for a cube is at least O(n) (where "n" is the number of voxels occupied in the cube) and has a large multiplication constant (consisting of neighbor candidate search, octant calculation, etc.). The solution complexity of a Wiener-Hopf system is O(t³) (where "t" is the number of filter tabs and t³ is less than or equal to n), and the multiplication constant is very small. The complexity of applying filters and coding / decoding filter coefficients is O(n), which also has a small multiplication constant. Therefore, it is important to identify the neighbors of a specific voxel in constant time (i.e., this means that filter tab calculations in the encoder / decoder must be performed in O(n) time).

[0317] The proposed acceleration algorithm extracts neighbor candidates from a set of adjacent 2kx2kx2k blocks that completely contain the rectangular neighbors of the processed voxels. A block (equivalent to a cube) corresponds to a subtree of an octree. Therefore, only the 3*k lower bits of the modal code of a voxel within a block change, while the upper bits remain fixed. We can assign block_code = (voxel_code >> 3*k), where voxel_code is the modal address of any voxel within the block and ">>" is a bitwise right shift. The key requirement is to access the list of voxels in each block in constant time. This is achieved through the following preprocessing steps.

[0318] 1) Store the occupied voxels of the slice in an array sorted by Morton code. All voxels of the block (cube) occupy contiguous slots in this array. The sorting complexity is O(NlogN), where N is the number of voxels occupied in the slice. For an appropriate N, since "logN" is much smaller than the multiplication constant in O(n) neighbor voxel search, we can assume O(n) complexity per cube.

[0319] 2) For each block, store the array1 position of the corresponding block in another array (array2) indexed by the block code.

[0320] Therefore, for every voxel, get the list of neighbors of the same block in constant time: array1[array2[voxel code >> 3*k]].

[0321] A block is defined by the voxel codes of an appropriately selected set of nearby voxels. An example is as follows (square neighbors are defined as ||neigh - vox|| < R).

[0322] a) Eight corner voxels of a cubic neighborhood (see Figure 6). The block size is selected such that k >= 2*R - 1 (to ensure the blocks are continuous).

[0323] b) 27 voxels: Corners + Middle Corners + Center (Rubik's Cube). K >= R - 1.

[0324] Option (b) requires exploring about twice as many candidate voxels as Option (a). However, for a suitable R neighborhood, there are only a few tens of voxels. Therefore, the overhead of experimentally selecting 27 - 8 = 19 additional blocks (including deduplication) overweights this twofold reduction in the number of voxels, and the computational cost of Option (a) is lower.

[0325] FIG. 15 shows an example of a 2D projection of a block defined by 8 corner phoxels and 27 corners + middle corners + center phoxels according to embodiments.

[0326] Memory requirements for neighbor voxel search according to embodiments:

[0327] Constructing array2 for the entire slice results in very large memory usage. The size of array2 is approximately equal to the number of blocks in the slice geometry (2kx2kx2k) and is not smaller than the number of blocks_codes between the block_code of the first block occupied in the slice and the block_code of the last block occupied. This number can be very large. However, this approach can slightly increase the compression ratio by using external cube taps for filtering.

[0328] FIG. 16 shows an example of retrieving truncated_block_code from a Morton address of a voxel according to embodiments.

[0329] The method according to the embodiments further includes a method of constructing array2 within a cube to reduce memory requirements and indexing elements using truncated_block_code. In this approach, the size of array2 is limited to 23*(pk), which is not large compared to the appropriate cube size. Only tabs within the cube can be used for filtering. The cube can be processed in parallel:

[0330] Truncated_block_code = (voxel_code & ((1 << 3*p) - 1)) >> 3*k,

[0331] Here, & is bitwise AND, << is bitwise left shift, and >> is bitwise right shift.

[0332] FIG. 17 shows information about a Molten code for neighbor search according to embodiments.

[0333] Neighbor voxel distribution for the corresponding filter tab according to the embodiments:

[0334] The neighbor voxel distribution for a filter tab according to the embodiments is processed according to the following principles: 1. distance, 2. direction

[0335] The method and apparatus according to the embodiments may further include and perform an acceleration of immediate neighbor voxel search operation.

[0336] The embodiments propose that the adjacent neighbors of a voxel are defined by the condition ||neigh - vox|| < R, where "neigh" and "vox" are vectors representing the respective voxel coordinates, and ||v|| represents the L2-norm of the vector 'v'. For R = 2, the adjacent neighbor detection problem is reduced to finding all adjacent neighbors (up to 26) of a voxel. This is a very important practical case because R = 2 is almost always the optimal choice for solid content filtering. Additionally, for R = 2 (and thus k = 1), the memory requirements of the basic search algorithm (23*(pk)) are substantial. The embodiments reduce computational complexity and memory usage compared to general search algorithms by introducing a special algorithm for adjacent neighbor search.

[0337] concept:

[0338] This algorithm processes voxels in the order of the Morton curve. If the Morton curve is regarded as a timeline, this idea can be explained as follows.

[0339] - For each voxel:

[0340] - Find "past" neighbors: Searches for previously saved voxels.

[0341] - Find neighbors of "current": Search voxels within the current 2x2x2 block.

[0342] - Do not search in the 'future' direction. Instead, save the voxels so they can be found in other voxels in the latter part of the curve.

[0343] This algorithm uses the following terms.

[0344] - Cube: A partition of size 2px2px2p corresponding to an octree subtree. 'p' is a constant of the ALF algorithm determined by the QP.

[0345] - Block: A 2kx2kx2k partition corresponding to an octree subtree. For integers 0 < k < p.

[0346] - Block size: Integer 'k' for a 2kx2kx2k block

[0347] - Border: One face of a 2x2x2 block (i.e., a square of size 2x2). Two adjacent blocks of size 1 share exactly one border (the two blocks are separated by that border). Note: A border can surround blocks of size 1, ���, and S < p (i.e., can be placed on one face).

[0348] - Boundary axis: Normal axis of the boundary plane (s, t, or v).

[0349] - Border Strength: The size of the largest block enclosed by the border. Note: Numerically, strength is equal to the number of consecutive 0s or 1s at less significant bit positions of coordinate values ​​along the boundary axis.

[0350] - Neighbors (for the current voxel): Voxels located in one of the 26 adjacent locations to the current voxel.

[0351] - Straight Neighbors: Neighbors located in adjacent 2x2x2 blocks

[0352] - Diagonal neighbors: Neighbors that are not in a straight line (i.e., located in a 2x2x2 block that is connected but not adjacent)

[0353] - Continuous Boundary: Voxels in adjacent 2x2x2 blocks crossing the boundary have a higher Morton code (i.e., follow on the Morton curve).

[0354] - Leading Boundary: Voxels in adjacent 2x2x2 blocks crossing the boundary have a lower Morton code (i.e., come ahead on the Morton curve). Note: For every voxel, each adjacent boundary is either continuous or leading.

[0355] - Inner Border (Block): All borders inside the block (therefore, strength is smaller than the block size)

[0356] - Outer Border (Block): All borders on the sides of the block (therefore, strength equal to the block size)

[0357] Algorithm Overview:

[0358] - For each voxel in the cube (voxels are processed in Morton curve order):

[0359] - Search all adjacent occupied boundaries (up to 3) in the storage.

[0360] - Searches for neighbors connected to the searched boundary.

[0361] - Search for neighbors located in the current 2x2x2 block (consider only voxels with a code higher than the current voxel's Morton code).

[0362] - For each neighbor

[0363] Add the current voxel and neighboring voxels to each other's tabs.

[0364] - Connect the current voxel to all (up to 3) adjacent boundaries.

[0365] - Add the subsequent boundary to the repository (if it does not already exist).

[0366] Each pair of voxels is examined / processed only once. The total complexity of traversing neighbors and calculating taps is O(n), with a very small constant factor (where 'n' is the number of occupied voxels in the cube). The quality, complexity, and memory requirements of the algorithm are essentially determined by two important aspects.

[0367] (1) Efficiency of storage operations: Indicates how efficiently storage, addition, and retrieval of subsequent / previous boundaries are organized.

[0368] (2) Types of voxel connections for boundaries: Not all pairs of neighboring voxels share a boundary. There are two main options for this algorithm.

[0369] - Identify only straight neighbors: Connect only the voxels of the current 2x2x2 block to the next boundary of the block. This approach skips up to 7 of the 26 potential neighbors for each voxel.

[0370] - Identify all neighbors: To support diagonal neighbors, additionally connect a voxel of one 2x2x2 block to the boundary of an adjacent 2x2x2 block. This mapping can be implemented efficiently.

[0371] Efficient storage operations: When the algorithm processes voxels along the Morton curve, there are two simple but important observations regarding boundary addition and search order.

[0372] 1) The internal boundaries of a block are processed before the first search operation for the next external boundary is executed (i.e., all addition and search operations are performed for all internal boundaries). Therefore, the idea of ​​preventing intersections by storing boundaries of different strengths separately is derived.

[0373] 2) The outer boundary of the block is added along the "inner" Morton curve (axes s, t, and v are connected in an intersecting order). The outer boundary is searched along the "outer" Morton curve (axis t is searched first, and then axes v and s are searched). Therefore, an idea is derived to prevent intersection by storing the boundaries of different axes separately.

[0374] In the embodiments, boundaries are stored in 3*(p-1) banks, one for each possible combination (boundary axis, boundary strength). The bank for a specific (axis, strength) combination is referred to as "storage[str_idx][ax_idx]", where str_idx = strength - 1 and ax_idx is 0 for "s", 1 for "t", and 2 for "v". Banks with lower ax_idx have larger memory capacity because they accumulate more boundaries before being reused / reset. Below is the frequency at which each bank can be reset / reused when voxel processing follows the Morton curve.

[0375] storage[str_idx][2]: every second block of size "strength".

[0376] storage[str_idx][1]: every fourth block of size "strength".

[0377] storage[str_idx][0]: the eighth block of size "strength" (i.e., each block of size "strength + 1").

[0378] The worst-case memory requirements for each bank are as follows.

[0379] Storage space[str_idx][2]: 2^(2* strength(strength) - 2)

[0380] Storage space[str_idx][1]: 2^(2*strength - 1)

[0381] Storage space[str_idx][0]: 2^(2*strength)

[0382] Executing all add / search operations within O(n) time (total) is essential for fast algorithm operations. This can be guaranteed when boundaries are stored in separate banks (as described above). The embodiments include two practical O(n) methods for addressing boundaries within the storage banks. For further explanation, it is assumed that boundaries are implemented as data structures containing an array of voxel identifiers (e.g., pointers or indices of Morton curves).

[0383] 1) Index mode.

[0384] Add / search operations access boundaries using the same index as the 2D Morton code. Projecting a 3D Morton curve onto a plane results in a 2D Morton curve. In this 2D projection, the Morton code of a voxel onto the boundary plane is used as an index to store the boundary adjacent to that voxel in the bank.

[0385] - Individual add / search operations are performed in O(1) time. The total add / search operation takes O(n) time.

[0386] - The memory requirement for this method is O(2(2*p)). By modifying the algorithm prescriptively to omit the boundary where strength >= max_strength, it can be reduced to O(2(2*max_strength)). The effect on ALF gain for max_strength = pm is negligible for small 'm' and sufficiently large cube size 'p'. That is, when the gain reduction is small, the memory requirement is reduced by 4 times, 16 times, etc.

[0387] - When resetting a bank for reuse, there is no need to reset / modify all O(2(2*strength)) boundary objects. Data structures and boundary structures may contain a "counter" that increases in the bank structure at each reset (O(1) complexity for bank reset). If the counters of the bank and boundary are different, the boundary slot is considered empty.

[0388] - Straight and diagonal neighbors can be connected to the boundary. A detailed explanation of diagonal connections is provided below.

[0389] - The constant factor in memory requirements increases the O(2(2*max_strength)) estimation time fourfold to support diagonal connections. Boundaries can be connected to up to 4 straight neighbors and up to 12 additional diagonal neighbors.

[0390] 2) FIFO (First-In, First-Out) mode.

[0391] The bank operates like a queue. The "add" operation adds a boundary to the end of the bank ('end' is updated). The "search" operation searches from the beginning of the bank until an adjacent boundary is found (or cannot be found). Any boundary not adjacent to the current voxel is immediately removed from the bank ('begin' is updated). The reason the FIFO operation mode works is that the order in which a Morton curve 'contacts' boundaries within the same bank (single axis, single intensity) is consistent with adjacent blocks of that size 'intensity'.

[0392] - Individual addition operations are performed in O(1) time. Individual search operations are performed in O(d) time. Here, 'd' is the number of un-deleted boundaries in the bank preceding the boundary adjacent to the current voxel. Therefore, the total time required for addition / search operations is O(n).

[0393] The memory requirement of this method is determined by the total number of occupied voxels adjacent to the boundary of strength p-1, and is largely independent of the voxel location within the cube. In the worst case (when all voxels are occupied), the memory requirement of this method is O(2(2*p)). However, on average, it is much smaller.

[0394] - O(1) complexity of bank reset

[0395] - Only straight neighbors can be connected to the boundary (a condition to prevent FIFO order violation).

[0396] Diagonal neighbor connection:

[0397] Straight neighbors are located in adjacent 2x2x2 blocks (i.e., blocks with different block codes along one axis). To identify straight neighbors, connect each voxel to an adjacent continuous boundary.

[0398] Diagonal neighbors are located in connected but non-adjacent 2x2x2 blocks (i.e., blocks with different block codes along two or three axes). To identify diagonal neighbors, the algorithm must map and connect the voxel to the contiguous boundaries of adjacent 2x2x2 blocks. To be eligible for diagonal mapping, there must be at least two contiguous boundaries adjacent to the voxel. This embodiment performs the mapping of the current voxel through the following steps.

[0399] - Select the smallest axis_idx among the continuous boundaries adjacent to the current voxel.

[0400] - Select a bank based on the selected boundary axis and the strength of adjacent boundaries along this axis.

[0401] - Select the bank boundaries (up to 3 target boundaries per bank) connected to the 2D projection of the current voxel (projected along the selected axis). The indices of these boundaries can be efficiently calculated using only the 3D Morton code of the current voxel and bit operations.

[0402] - Connects the current voxel to the selected border of the selected bank.

[0403] - Repeat the above 4 steps for the second smallest axis_idx.

[0404] - Prevents duplicate connections when border strengths are the same along the current and previous axis_idx.

[0405] - If there are three consecutive borders adjacent to the current voxel

[0406] - Repeat step 4 above for the third axis_idx.

[0407] - Prevents duplicate connections when the border strength along the current axis_idx is equal to the border strength along one of the previous two axis_idxes.

[0408] Consequently, all neighbors adjacent to the current voxel in straight and diagonal directions in each direction (defined as the boundary axis) are connected to the current voxel's adjacent previous boundary. All neighbors are identified by the diagonal connection algorithm.

[0409] Skip neighbor

[0410] Neighbor search algorithms can significantly reduce computational complexity and memory usage by omitting some neighbors. It should be noted that the ALF design of this embodiment exhibits robustness against omitted neighbors because taps are calculated by averaging across the entire volume.

[0411] In the previous chapter, two practical methods for skipping neighbors for immediate neighbor discovery were presented.

[0412] - Skip diagonal neighbors to avoid complexity caused by diagonal connections and additional memory for storing diagonal connections.

[0413] - Skip the highest strength boundary in the cube to reduce memory usage for 4x. Here, 'x' is the number of skipped strengths.

[0414] The skipping method can be generalized to larger window sizes. For example, if the current voxel is located within 2kx2kx2k, 2kx2kx2k corner blocks with smaller intersections with the current voxel block's spherical neighbors can be discarded without significantly reducing the ALF gain. Such modifications to the search algorithm are standard. It has been proposed to add corresponding syntax elements to APS to support various types of search optimization.

[0415] FIGS. 18, 19, and 20 show filter window sizes according to embodiments.

[0416] Determining the Optimal Filter Window Size

[0417] Determining a window size suitable for a given point cloud is critical to ALF performance. The present invention includes a method for selecting an optimal window. This method is based on voxel density within the cloud. This method can be utilized symmetrically on both the encoder and decoder sides. In this case, no bits are consumed for transmitting the window size. A new flag indicating whether a window size syntax element is present in the ADU header has been added to the APS.

[0418] A dataset containing the optimal window size for point cloud sequences was constructed from the GPCC CTC sequence set. The optimal window for each sequence was determined by calculating all acceptable window sizes. The empirical dataset showed that the window size depends significantly on the voxel density within the sequence. The average number of voxels within a cube (average per sequence, for filled cubes only, defined by a fixed threshold for the minimum number of voxels within the cube) is a sustainable criterion for determining the optimal window.

[0419] Figure 18 shows the dependence of the optimal window size on the average (per sequence) number of phoxels in a 28x28x28 cube, threshold = 500.

[0420] The discovered dependency is well approximated by a hyperbolic equation with two parameters win_size = a / (#voxels)+b. The optimal parameters a and b were derived based on experimental data (a = 60000, b = 2).

[0421] Figure 19 shows the optimal window size approximation by a / (#voxels)+b, a = 60000, b = 2.

[0422] FIG. 20 shows the optimal window size approximation according to floor(a / (#voxels)+b+0.5), a = 60000, b = 2.

[0423] This formula can be used to determine the optimal window size for sequences, slices, or individual cubes.

[0424] Distribution of adjacent voxels for the corresponding filter tab.

[0425] In this embodiment, the distribution of adjacent voxels for the filter tab is processed according to the following principle.

[0426] 1. Distance

[0427] 2. Direction

[0428] FIG. 21 shows the ploxel distribution of filter tabs according to distance (2D and 3D cases) according to embodiments.

[0429] 1. Distance:

[0430] According to this rule, the attribute values ​​of the reconstructed voxels are distributed to different filter tabs based on the distance between neighboring voxels and the center (processed voxel).

[0431] Figure 21 shows three regions for neighboring voxel distributions. In this example, the four tab filters can be thought of as follows.

[0432] Filter = A1value1 + A2value2 + A3value3 + A4

[0433] value1 is the attribute value of the processed voxel, and value2 is calculated using all voxel attributes inside the small sphere.

[0434] value3 is calculated using all voxel properties inside the large sphere and outside the small sphere, and A4 is the offset. There is no guarantee that voxels exist in the calculated value1 and value2, in which case the default value is 0.

[0435] The distance between voxels can be any function that defines a metric in 3D space.

[0436] e.g.: abs(dx)+ abs(dy)+ abs(dz) or sqrt(dx2 + dy2+dz2)

[0437] FIG. 22 shows the shape of a filter according to the embodiments.

[0438] 2. Direction:

[0439] direction.

[0440] To better adapt to the local shape of a given point cloud, different tabs can be assigned to different directions. However, the development of ALF for modern video codecs presents simplified assumptions regarding the behavior of the autocorrelation function. This function is symmetric with respect to the center.

[0441] FIG. 23 shows the phoxel distribution (3D case) of filter tabs according to the direction according to the embodiments.

[0442] This leads to the clear conclusion that the filter coefficients for opposite points are identical and the overhead can be reduced by half. According to this rule, the attribute values ​​of the reconstructed voxels are distributed across different filter tabs depending on the direction from the center (processed voxel) to neighboring voxels, but voxels with opposite directions are integrated into the same region. The distribution of the eight octants (directions) for neighboring voxels is shown in Fig. 22. Symmetrical octants are integrated into a single region.

[0443] In this example, six tab filters can be considered.

[0444] Filter = A0value0 + A1value1 + A2value2 + A3 value3 + A4 value4 + A5

[0445] Here, value0 is the attribute value of the processed voxel. value1, value2, value3, value4, and value5 are calculated using all voxel attributes within each octant, and A5 is the offset.

[0446] FIG. 24 shows an example of a 4-tab symmetric ALF design using 8 octants according to embodiments.

[0447] When the attribute values ​​of neighboring voxels are equal to or very close to those of the processed voxel, the condition number of the matrix in the Wiener-Hopf equation can be high. This means that the problem becomes an ill-posed problem. When calculating the value, it is better to calculate the difference in attributes between the neighboring voxel and the processed voxel instead of using the attributes. This technique improves the situation when the solution to the Wiener-Hopf equation is an integral problem and significantly reduces the dynamic range of intermediate calculations. Assume that Y0 is the attribute value of the processed voxel and Yk (k=1, N1) is the attribute value of the neighboring voxel belonging to two symmetric octants 1 and 5 (Fig. 21). Therefore, the value 1 is calculated using the set Yk - Y0. To determine the value 1, the median of {Yk - Y0} or the most frequently occurring value in the dataset can be used.

[0448] value 1=med{Y_k-Y_0} or value1=mode{Y_k-Y_0}

[0449] When Yk-Y0 is an integer value, the median and mode are also integers.

[0450] Another possibility is the average value. This can be calculated very easily, but when Yk-Y0 is an integer, the average value is very likely not an integer.

[0451] Apply Filter:

[0452] Filter = A0value0+A1value1+ A2value2+A3value3+A4value4+ A5

[0453] The results found using the difference between value1, value2, value1, and value1 show that in most cases, A0 is very close to "1" and A5 is very close to "0". This helps in designing most simple 4-tap filters.

[0454] Filter = value0+A1value1+ A2value2+A3value3+A4value4

[0455] The process of calculating the matrix and dominance column vector of the Wiener-Hopf equation using the original and reconstructed attribute signals according to the embodiments is as follows:

[0456] Without compromising generality, the filter takes the following form. Even if an offset is required, the last value N It is equal to 1.

[0457]

[0458] The unknown coefficient {coeffn} can be derived from the minimum value of the squared error.

[0459]

[0460] The problem in this formula is linear regression. The necessary condition for obtaining the minimum value is derived by differentiating with respect to the coefficients of the unknown, which leads to a linear system known as the Wiener-Hoff equation.

[0461]

[0462]

[0463]

[0464]

[0465] In most cases, the matrix is ​​symmetric and positive definite. However, if the determinant of matrix (A) is 0, the problem becomes uncertain, making ridge regression a more desirable approach.

[0466]

[0467] Here, ε>0 is a small positive number and I is the same matrix. This not only solves the integration problem but also reduces the vector length and the number of bits required for the coefficients.

[0468] Solution to a system of linear equations (see reference). This is an encoder-only option. While Gaussian elimination is always possible, Cholesky decomposition is a much faster and more stable alternative due to the symmetric and positive definite properties of matrices. This technique is also known as the square root method.

[0469] The Kolleski decomposition for a 3x3 matrix is ​​given by the following formula.

[0470]

[0471] Scaling factor and rounding operation according to embodiments:

[0472] Numerous experiments have shown that the coefficient value1 is close to "1" and the offset is close to "0". Therefore, these values ​​are used as defaults to reduce overhead. To reduce correlation, deltak = (valuek - value0) is used.

[0473] The final design is as follows.

[0474]

[0475] Matrix A and B are as follows:

[0476]

[0477] Filter coefficients are derived by solving the Wiener-Hopf equation. The resulting filter coefficients are in floating-point format, but representing variables as floating-point values ​​is particularly difficult in hardware design. If the fractional part of the filter coefficients is limited to 2k bits, the derived filter coefficients must be scaled and rounded to this precision. The scaling factor QSTEP was selected as 2k = 2, 4, 8, 16, 32, etc., to allow the use of shift operations instead of division.

[0478] FIG. 25 shows a flowchart of scaling and rounding according to embodiments.

[0479] Filter coefficient limitation according to embodiments:

[0480] Various experimental results show that most coefficients before scaling fall within the [-0.5, 0.5] range. Filter coefficient limits were introduced to prevent integer overflow and reduce overhead. On the encoder side, limits on the absolute values ​​of the coefficients are required. This could involve naive clipping or using a different solution (A+εI)C=B with a higher ε. If it is known in advance that all unscaled coefficients fall within the [-0.5, 0.5] range, then integer coefficients fall within the [-QSTEP / 2, QSTEP / 2] range. With this prior information, overhead can be significantly reduced by using a truncated unary code or a Rice code.

[0481] Calculation of convolution filter coefficients using attribute values ​​of referenced voxels according to embodiments:

[0482] The application of adaptive loop filters to each voxel is independent and can be performed in parallel. Assume that QSTEP is selected as 2k.

[0483] And Icoeff is the scaled and rounded coefficient corresponding to the coefficient.

[0484]

[0485] Since Qstep is a power of 2, the final operation " / " can be performed using a shift operation (>>k). Clipping operations ensure that attribute values ​​do not exceed the allowed range. QSTEP coefficients can be -1, 0, or 1. This means that filters without multiplication can be designed.

[0486] FIG. 26 shows a flowchart of the calculation of ALF tap values ​​according to embodiments.

[0487] ALF process using average values ​​in integer operations according to embodiments:

[0488] The input to this process is an array of differences between the processed voxels and neighboring voxels prior to the adaptive loop filter DELTA_L. The output of this process is an array of reconstructed image samples modified after the adaptive loop filter alfFilterL.

[0489]

[0490]

[0491] For example, N of 0 to 10 k for In the case of, in the above formula is a lookup table mult(N k It can be replaced with ).

[0492] mult

[0011] = { 0, 1024, 512, 341, 256, 204, 170 , 146, 128, 113, 102}

[0493] Therefore, in the ALF process, it is necessary to store the sum of attribute values ​​and the number of adjacent ploxes.

[0494] FIG. 27 shows the dynamic range for calculating a four-tap filter according to the embodiments.

[0495] Dynamic range of 4-tap filter calculation using average values ​​in integer operations according to embodiments:

[0496] QSTEP k Selected as Assuming the attribute is [0, 2m-1], all calculations correspond to 2+k+n+m bits. For example, QSTEP=64, k=6, 2^n=1024, n=10, the attribute is [0, 255], and m=8. All calculations correspond to a 26-bit buffer.

[0497] Application determination based on RDO or Saliency map according to embodiments:

[0498] To improve the coding efficiency of ALF applications, it can be determined through Rate Distortion Optimization.

[0499] In terms of encoding, SSD norms can be calculated in advance before filtering.

[0500]

[0501] After filtration,

[0502]

[0503] The ratio distortion optimization decision can be performed based on the following simple conditions.

[0504]

[0505] Here, "bits" refers to the number of bits required for the overhead of the current cube. This number of bits can be estimated with high accuracy using a real-world model or approximated using only integer AlF coefficients. If the condition is true, it means that additional overhead is required for quality improvement and a filter must be applied. Parameters The optimal value of follows the following law.

[0506]

[0507] However, this constant varies depending on the color component, slice type, scarcity, etc.

[0508] FIG. 28 illustrates an RDO-based application of ALF for various QPs according to embodiments.

[0509] As shown in Fig. 28, it can be seen that ALF is applied to the target regions depending on the activation of ALF.

[0510] The subjective improvement of ALF application can be verified through the determination of the salience map. It is well known that viewers do not pay attention to the entire image, but rather focus only on small, specific parts.

[0511] FIG. 29 shows an example of the most interesting part in the image according to the embodiments.

[0512] Saliency maps can be found by the encoder using algorithms or provided by the user. ALF contributes more to subjective quality by reducing coding errors in saliency areas.

[0513] The encoding method according to the embodiments can encode point cloud data and generate related parameters (which may be referred to as signaling information, metadata, etc.). The encoding method according to the embodiments can generate a bitstream including point cloud data and parameters. The decoding method according to the embodiments can decode point cloud data within the bitstream based on parameters within the bitstream.

[0514] High-level syntax flag

[0515] In the process of encoding / decoding attribute information, syntax related to the Adaptive Loop Filter (ALF) can be added to the Attribute Parameter Set and Attribute Data Unit for signaling.

[0516] FIG. 30 shows the APS (attribute parameter set) syntax in the bitstream.

[0517] The encoding method according to the embodiments (transmitting device (10000) of FIG. 1, point cloud video encoder (10002), transmitter (10003), acquisition-encoding-transmission (20000-20001-20002) of FIG. 2, encoding based on FIG. 11 to 29, generation of syntax information in FIG. 30 to 32, filtering in FIG. 33 to 35, generation of syntax information in FIG. 36a and FIG. 36b to 37, encoder in FIG. 38, encoding method in FIG. 40) can generate parameter information as in FIG. 30.

[0518] A decoding method according to embodiments (receiving device (10004) of FIG. 1, receiver (10005), point cloud video decoder (10006), transmission-decoding-rendering (20002-20003-20004) of FIG. 2, decoder of FIG. 7, receiving device of FIG. 9, device of FIG. 10, encoding based on FIG. 11 to 29, acquisition of syntax information of FIG. 30 to 32, filtering of FIG. 33 to 35, acquisition of syntax information of FIG. 36a and FIG. 36b to 37, encoder of FIG. 39, encoding method of FIG. 41) can decode point cloud data by acquiring parameter information from a bitstream as in FIG. 30.

[0519] Parameters specified by the APS apply to all DUs where the APS is enabled.

[0520] aps_attr_parameter_set_id identifies the APS so that other DUs can reference it.

[0521] aps_seq_parameter_set_id identifies the active SPS as sps_seq_parameter_set_id.

[0522] attr_coding_type specifies the attribute coding method. Valid values ​​are specified in Table X. Other values ​​are reserved by ISO / IEC for future use. Decoders complying with this version of this document must ignore (remove from the bitstream and discard) attribute data units coded with the reserved values ​​of attr_coding_type.

[0523] attr_coding_type:

[0524]

[0525] attr_primary_qp_minus4 + 4 represents the QP of the primary attribute component before adding offsets per slice, per region, and per transform level.

[0526] attr_secondary_qp_offset represents the offset to be applied to the primary attribute QP to derive the QPs of all secondary attribute components.

[0527] attr_qp_offsets_present indicates whether attr_qp_offset[ ], which is the attribute QP offset per slice, is present in the ADU header (if 1) or not (if 0).

[0528] aps_extension_present indicates whether the aps_extension_data syntax element is present in the APS syntax structure. For bitstreams conforming to this version of this document, aps_extension_present must be 0. The value of aps_extension_present 1 is reserved for future use by ISO / IEC.

[0529] aps_extension_data can have any value. The presence or value of this parameter does not affect the decoder's suitability for the profiles specified in this version of this document. The decoder must ignore all aps_extension_data syntax elements.

[0530] attr_adaptive_loop_filter_enabled indicates whether to apply attribute coding using an adaptive loop filter (if 1) or not (if 0). If attr_adaptive_loop_filter_enabled is not present, it is inferred to be 0.

[0531] attr_alf_component_type indicates the attribute component to which the adaptive loop filter is applied. Valid values ​​are specified in the table.

[0532] attr_alf_component_type:

[0533]

[0534] attr_alf_neighbor_search_type indicates the algorithm used for neighbor voxel search. Valid values ​​are specified in the table.

[0535] attr_alf_neighbor_search_type:

[0536]

[0537] attr_alf_window_size_present_flag indicates whether the attr_alf_window_size syntax element is present in the ADU header.

[0538] FIG. 31 shows an attribute data unit according to embodiments.

[0539] An ADU codes attribute values ​​for a single attribute within a slice. An ADU consists of an ADU header and attribute coefficients (attribute_coeffs) (if transform coding is enabled) or directly coded attribute values ​​(attribute_raw).

[0540] When transform coding and adaptive loop filters are enabled, attribute ALF coefficients (attribute_alf_coeffs) consist of an ADU header and attribute coefficients.

[0541] FIG. 32 shows an attribute data unit header according to embodiments.

[0542] adu_attr_parameter_set_id represents the active APS as aps_attr_parameter_set_id.

[0543] adu_temporal_id represents the time ID of the frame associated with the attribute data unit.

[0544] adu_sps_attr_idx identifies the attribute coded as an index in the list of active SPS attributes. At the beginning of every ADU, the variable AttrIdx is set to adu_sps_attr_idx.

[0545] AttrIdx = adu_sps_attr_idx

[0546] If attr_coding_type is not 3, the attribute coded with ADU has a maximum of 3 components.

[0547] adu_slice_id represents the slice_id value of the previous GDU.

[0548] attr_alf_window_size represents the ALF window size.

[0549] attr_alf_qstep_log2 represents the ALF coefficient quantization step.

[0550] attr_alf_geom_shift represents the geometric coordinate shift value for ALF.

[0551] Predict filter coefficients from already processed blocks.

[0552] The coefficients of adjacent regions show a high correlation. In particular, in the 70% case, the coefficients improve the quality of not only the current region but also the next region. Therefore, filter coefficient prediction can reduce overhead.

[0553] Region-based ALF for chroma components.

[0554] Adaptive loop filters can be applied only to the luminance component because luminance quality is more important. This is based on the fact that human vision is more sensitive to changes in luminance than to changes in color

[0015] . This is one of the reasons why chroma subsumping is commonly used. However, the results of adaptive loop filters bring not only quality improvement (in the case of intra) but also bit reduction (in the case of inter). Applying ALF to the current frame improves the quality of the reference frame and reduces motion compensation attenuation overhead. Therefore, using ALF on intra results in only a slight increase in bit count, while it provides actual bit savings for inter. Users can attempt additional bit savings by applying ALF to the chroma component. There are three possible application methods: Chroma by Chroma, Chroma by Luma, Chroma by Luma, Chroma by Luma, Chroma by Luma.

[0555] 1. ALF Chroma by Chroma. This process is generally similar to Luma ALF. However, as mentioned above, since the Luma component is more important, it is reasonable to apply the Chroma filter after Luma ALF. This can reduce the complexity of adjacent voxel search. However, a new Wiener-Hopf equation must be solved to determine the filter coefficients, and the final filtering is as follows.

[0556]

[0557] Here, value0 and delta k is derived from the chroma component of a neighboring volcel.

[0558] ALF chroma using Luma. This process uses the same delta as already used in Luma. kIt uses . That is, the matrix of the Wiener-Hopf equation is identical, meaning that the inverse matrix has already been found when applying ALF to the lumens. Only the column vectors on the right-hand side of the Wiener-Hopf equation need to be recalculated. To determine the filter coefficients, the Wiener-Hopf equation, for which the inverse matrix is ​​already known, must be solved. The final filtering is as follows.

[0559]

[0560] Here, value0 is derived from the chroma component, but delta k It is borrowed from the Ruma component.

[0561] ALF chroma uses both luma and chroma. This process is recommended only when the number of bits used for chroma encoding is significantly higher than the codec used for luma.

[0562] Truncated rice codes according to the embodiments:

[0563] Truncated rice codes are used to code the absolute values ​​of filter coefficients. Geometric distribution For , the optimal rice parameter n can be obtained using the following equation.

[0564]

[0565] Scaling factor When is ultimately considered

[0566]

[0567]

[0568] This formula roughly estimates the Rice parameter by assuming a geometric distribution and bypassing coding. However, it explains why the Rice parameter must increase as the scaling factor increases.

[0569] Entropy coding is used to handle encoding overhead (integer ALF coefficients, ALF flags, etc.). The encoder / decoder side already knows how many cubes are used for each slice / frame. First, the encoder determines whether it is necessary to write an ALF on / off flag to each cube. If the encoder determines that this flag is too expensive, the coefficient is recorded as 0 when ALF is not applied to the current cube.

[0570] Overhead for Luma according to the embodiments (integer ALF coefficient):

[0571] The attribute filter coefficient encoding pseudocode is as follows:

[0572] void encodeCoeffALF (vector <int>& flag, vector <int>& coeff,PCCResidualsEncoder& encoder,bool filter_apply) {if(filter_apply == false) return;int n = flag.size();int index_coef = 0;bool need_to_write_flag=check_need_to_write_flag(nuber_of_zeros, n);for (int i = 0; i < n; i++){if (need_to_write_flag) { encodeALFflag(flag[i]);if (flag[i]) encodecoeffLast());}else {if (flag[i]){ encodecoeff();} else {encodezeros();}

[0573] FIG. 33 shows a flowchart of ALF coefficient coding according to embodiments. The method according to embodiments is as follows: start the encodeCoeffALF function. check if filter_apply is false. If it is false, the filter is not applied, so terminate. check the number of zeros. If a filter needs to be applied, call the check_need_flag(number_of_zeros) function. This determines whether special processing is required by identifying the number of zero coefficients. It repeats as long as index i is less than n and branches into two paths. 1) Zero coefficient processing: If flag is false, call the encodezeros function to efficiently encode consecutive zero coefficients. Then, if flag becomes true, call the encodecoeff function to encode the actual coefficient values. If i >= n, terminate the processing. 2) ALF flag processing: If need_flag is true and i < n, process the ALF flag. If flag is false, terminate immediately. If flag is true, call the encodecoeffLast function to specially encode the last coefficient.

[0574] FIG. 34 shows a flowchart of ALF coefficient coding for four tap filters encodeCoeff, encodeCoeffLast, and encodeCoeffzeros according to embodiments.

[0575] Coeff Encoding Process: Sequentially encode the 4 coefficients from coeff[index + 0] to coeff[index + 3]. For each coefficient, call the encodeALFcoeffRice() function to encode using the Rice coding method. After processing all coefficients, increment the index by 4 (index = index + 4).

[0576] CoeffLast Encoding Process: Encode the first three coefficients (coeff[index + 0], coeff[index + 1], coeff[index + 2]) normally. Check the needLast flag at the fourth coefficient (coeff[index + 3]). If needLast is true: Encode the last coefficient in a special way using the RiceLast() function. If needLast is false: Use the encodeALFcoeffRice() function. After processing is complete, increment the index by 4.

[0577] encodeZeros process: This is a function designed to efficiently process consecutive zero counts. It calls encodeALFcoeffRice(0) four times in a row. This process encodes that all counts are zero. Instead of encoding zero values ​​individually, they are processed as a group to increase compression efficiency.

[0578] If the flag is set and the first three coefficients are 0, the last coefficient is definitely not 0, so encodeCoeffLast can be applied to reduce the absolute value of the coefficient by 1. If the flag is not applied, encodezeros is used.

[0579] For the 4-tap filter, nine context models are required for overhead encoding. One context model ctxALF[0] for flags, four context models ctxALF[1], ctxALF[2], ctxALF[3], ctxALF[4] for absolute values, and four context models ctxALF[5], ctxALF[6], ctxALF[7], ctxALF[8] for the sign of each coefficient are required.

[0580] The truncated Rice coding for the ALF coefficient pseudocode is as follows:

[0581] void encodeALFcoeffRice (int coeff, int index, int max, int Riceparameter, arithmeticEncoder) {

[0582] Level =abs(coeff);

[0583] Level1= Level>>Riceparameter; int max1 = max >> Riceparameter;

[0584] max1 = max >> Riceparameter; reminder = level - (level1 << Riceparameter)

[0585] reminder = level - (level1 << Riceparameter);

[0586] for ( i = 0; i < level1; i++){

[0587] arithmeticEncoder.encode(0, ctx[index+1]);

[0588] if (level1 < max1 - 1)

[0589] arithmeticEncoder.encode(1, ctx[index+1]);

[0590] for ( i = 0; i < Riceparameter; i++){

[0591] simbol = reminder & 1;

[0592] arithmeticEncoder.encode_by_pass(simbol);

[0593] if (level > 0){

[0594] reminder =reminder >>1;}

[0595] if (level > 0){

[0596] arithmeticEncoder.encode(sign, ctx[5+ index]);}

[0597] FIG. 35 shows a flowchart of an ALF coefficient decoding process according to embodiments.

[0598] The method of FIG. 35 is the ALF coefficient decoding of the decodeCoeffALF function and operates in the reverse order of the encoding process of FIG. 33 to FIG. 34:

[0599] The decodeCoeffALF function starts. First, the decodeALFflag() function is called to decode whether a filter is applied. If filter_apply is false, the filter is not used, so the process terminates immediately. The flag information is decoded. If a filter is applied, the decodeALFflag() function is called again to decode additional flag information. The need_flag value is checked. 1) Count Decoding: Follow this path if need_flag is false. Repeat as long as the condition i < sizeflag is satisfied. The decodeALF() and Checflag() functions are called to decode the counts. Processing continues by incrementing index i. 2) Flag-Based Decoding: Follow this path if need_flag is true. Under the condition i < sizeflag: Call the decodeALFflag() function to decode the flag at each position. Check the decoded flag value. If flag is true: Call the decodeALFLast() function to specifically decode the last count. If flag is false: Skip the corresponding coefficient.

[0600] The syntax and semantics regarding encoding and decoding according to the embodiments are as follows. The attribute data unit according to the embodiments further includes syntax element(s) for the coded ALF coefficients.

[0601] FIGS. 36a and FIGS. 36b show attribute ALF coefficients according to the embodiments.

[0602] FIG. 37 shows an attribute ALF coefficient tuple according to the embodiments.

[0603] Attribute data unit semantics

[0604] An ADU codes attribute values ​​for a single attribute within a slice. An ADU consists of an ADU header, attribute coefficients (attribute_coeffs) or directly coded attribute values ​​(attribute_raw) if transform coding is enabled, and filter coefficients within a loop if filtering is enabled in the ADU header.

[0605] Adds semantics for the ALF coefficients coded in the semantics of the sub-clause syntax elements of the slice attribute clause.

[0606] Attribute data unit ALF coefficient

[0607] The AttrALFCoeff array has elements AttrALFCoeff[cubeidx ][tapsidx] and contains scaled filter coefficient values. All elements of the array must be initialized to 0. AttrALFCoeff contains NumALFTaps filter coefficients per attribute channel per cube. The first NumALFTaps-1 coefficients are coded in attribute_alf_coeff_tuple(). The last filter coefficient is for the center (filtered) voxel and is inferred as 1. If all filter coefficients coded in attribute_alf_coeff_tuple() are 0, ALF is not applied to the c-th channel of the cube at the corresponding cubeidx.

[0608] adu_alf_flag indicates whether the ADU has ALF filter coefficients (i.e., whether ALF is applied to the reconstructed voxels belonging to that ADU).

[0609] If alf_flag_cube_present is 1, the application of filters to each cube is explicitly signaled by flags alf_flag_cube, alf_flag_cube_cb, and alf_flag_cube_cr. If alf_flag_cube_present is 0, ALF coefficients for each cube are transmitted.

[0610] Attribute ALF Count Tuple

[0611] According to zeroCoeffs, alf_coeff_abs[tapsidx], last_alf_coeff_abs_minus1, and alf_coeff_sign[tapsidx] together represent the tapsidx-th filter coefficient ALFAttrCoeff[cubeidx][c]. coeff_sign[c] specifies whether the sign of the coefficient is positive (when 0) or negative (when 1). zeroCoeffs is equal to 1 when the filter coefficients of the tuple are 0, ⪋, and NumALFTaps-2.

[0612] The coefficients of the cubeidx-th cube and the nth attribute component are specified through the derivation of ALFAttrCoeff.

[0613] Zero factor = 1

[0614] for (tapsidx = 0; tapsidx < NumALFTaps - 1; tapsidx++){

[0615] absVal = alf_coeff_abs[tapsidx]

[0616] if (absVal)

[0617] Zero factor = 0

[0618] if (tapsidx == NumALFTapsMinus1- 1 && zeroCoeffs)

[0619] absVal = last_alf_coeff_abs_minus1 + 1

[0620] ALFAttrCoeff[cubeidx][c][tapsidx] = (1 - 2 Х alf_coeff_sign[tapsidx]) Х absVal

[0621] ALFAttrCoeff[cubeidx][c][NumALFTaps - 1] = 1

[0622] Add property filtering using an adaptive loop filter to the slice property.

[0623] Attribute filtering using adaptive loop filters

[0624] common

[0625] The Adaptive Loop Filter (ALF) is an FIR in-loop filtering technique that aims to minimize the Mean Squared Element (MSE) distance between the reconstructed attribute plane and the original attribute plane. This filter is applied when attr_alf_enabled is 1. The encoder calculates filter coefficients by solving the Wiener-Hopf equation using floating-point operations. The decoder reads the filter coefficients / kernel from the ADU, calculates the filter tap on the reconstructed attribute plane, and performs convolution between the calculated tap and the filter kernel; all of these processes use integer operations. The filtering process takes the reconstructed attribute coefficient PointAttr as input and outputs the filtered attribute coefficient FilteredPointAttr.

[0626] Filtering preparation

[0627] The filtering process is specified for a single voxel. The filtering process begins with a preparation phase consisting of the following.

[0628] Read the ALF coefficient quantization step as follows: ALFQstep = Exp2(attr_alf_qstep_log2), filter window size ALFWin =attr_alf_window_size, geometry coordinate shift = ALFGeomShift = attr_alf_geom_shift.

[0629] The voxel coordinates are moved to adjust the size of the geometry. This movement may generate additional duplicate points, but it does not affect the filter application process.

[0630] for (ptIdx = 0; ptIdx < PointCnt; ptIdx++)

[0631] for (k = 0; k < 3; k++) {

[0632] ALFAttrPos[ptIdx][k] = AttrPos[ptIdx][k] >> ALFGeomShift}

[0633] Determine cube size (log, ALFLog2CubeSize): ALFLog2CubeSize := 9 + (qpP - 34) / 6

[0634] Calculate the number of filled cubes, NumPopulatedCubes, by iterating through all cubes in the slice (cube size is 1 << ALFLog2CubeSize) and counting the number of cubes containing more voxels than ALFCubeVoxelThreshold.

[0635] Filter coefficient order

[0636] The filter is applied if NumPopulatedCubes > 0 and alf_flag == 1. Then, the filter coefficients are decoded (the ALFAttrCoeff array is created). ALFAttrCoeff contains only the filter coefficients for the filled cubes. cubeidx enumerates the filled cubes in ascending order of ALFCubeCode.

[0637] ALFCubeCode = Morton[ALFAttrPos[cubePtIdx]] >> 3 * ALFLog2CubeSize

[0638] Here, cubePtIdx is the index of all foxholes belonging to the cube.

[0639] Attribute Filtering

[0640] Filtering is applied independently to each cube. Filter tabs in adjacent cubes must be removed from the decoder.

[0641] Filter tab calculation

[0642] The filter tab is calculated using the attribute values ​​of voxels in the voxel's rectangular neighbors. For the current voxel where ptIdx = curPtIdx, neighbor voxel candidates where ptIdx = candPtIdx must satisfy the following conditions.

[0643] (ExSquared[ALFAttrPos[curPtIdx][0] - ALFAttrPos[candPtIdx][0]] + ExSquared[ALFAttrPos[curPtIdx][1] - ALFAttrPos[candPtIdx][1]] + ExSquared[ALFAttrPos[curPtIdx][2] - ALFAttrPos[candPtIdx][2]]) < ExSquared[ALFWin]

[0644] To calculate the filter tab, the decoder must identify all these adjacent voxels.

[0645] The current filter design implies NumALFTaps = 5. Taps with tapsidx = 0, 1, 2, 3 are defined as 8 octants of spherical neighbors. Voxels of opposite octants contribute to the same tap. For each adjacent voxel, a tap index is assigned using the following logic.

[0646] dk0Gt0 = ALFAttrPos[curPtIdx][0] - ALFAttrPos[candPtIdx][0]

[0647] dk1Gt0 = ALFAttrPos[curPtIdx][1] - ALFAttrPos[candPtIdx][1]

[0648] dk2Gt0 = ALFAttrPos[curPtIdx][2] - ALFAttrPos[candPtIdx][2]

[0649] if ((dk0Gt0 == true && dk1Gt0 == true && dk2Gt0 == true) || (dk0Gt0 == false && dk1Gt0 == false && dk2Gt0 == false))

[0650] tapsidx = 0;

[0651] else if ((dk0Gt0 == false && dk1Gt0 == true && dk2Gt0 == true) || (dk0Gt0 == true && dk1Gt0 == false && dk2Gt0 == false))

[0652] tapsidx = 1;

[0653] else if ((dk0Gt0 == false && dk1Gt0 == false && dk2Gt0 == true) || (dk0Gt0 == true && dk1Gt0 == true && dk2Gt0 == false))

[0654] tapsidx = 3;

[0655] else

[0656] tapsidx = 4;

[0657] The filter tab for the c-th attribute of the curPtIdx-th voxel can be explained through the derivation of ALFTaps. When tapsidx = 0, ..., 3, ALFTaps[curPtIdx][c][tapsidx] is equal to the difference between the attribute value of the current voxel and the arithmetic mean of the attribute values ​​of neighboring voxels with tap index tapsidx. Also, ALFTaps[curPtIdx][c][NumALFTaps - 1] = PointAttr[curPtIdx][c]. Floating-point operations are required for the ALFTaps expression.

[0658] The decoder performs only integer calculations. Therefore, the filter tab for the c-th attribute of the curPtIdx-th voxel is specified through the derivation of ALFTapsSumAttributes and ALFTapsNumPt. Here, ALFTapsSumAttributes[c][tapsidx] is the sum of the attribute values ​​of neighboring voxels with tap index tapsidx, and ALFTapsNumPt[tapsidx] (same for all c) is the number of such neighboring voxels. The calculation of the arithmetic mean is integrated into the filter coefficient scaling process, where division is replaced by multiplication and shifting.

[0659] Apply filter

[0660] The inputs to this process are the reconstructed attribute coefficients PointAttr[curPtIdx][c], the filter tabs ALFTapsSumAttributes[c] and ALFTapsNumPt, the decoded filter coefficients ALFAttrCoeff[cubeidx][c], and the scaling factor ALFQstep. The output is the filtered reconstructed coefficients FilteredPointAttr[curPtIdx][c].

[0661] The decoded filter coefficient ALFAttrCoeff is a scaled version of the final filter coefficient with a scaling factor ALFQstep. The effect of the division AttrALFCoeff[cubeidx ][c][tapsidx] / ALFQstep in the coefficient calculation and the effect of the division ALFTapsSumAttributes[c][tapsidx] / ALFTapsNumPt[tapsidx] in the tap calculation are combined into a single integer value calculation. ALFQstep is equal to a power of 2.

[0662] Fixed-point operations using bit precision ALFBitPrecision are used.

[0663] Integer array ALFMult:

[0664] const int ALFMult

[0101] = {

[0665] 0, 1024, 512, 341, 256, 204, 170, 146, 128, 113, 102, 93, 85, 78, 73,

[0666] 68, 64, 60, 56, 53, 51, 48, 46, 44, 42, 40, 39, 37, 36, 35,

[0667] 34, 33, 32, 31, 30, 29, 28, 27, 26, 26, 25, 24, 24, 23, 23,

[0668] 22, 22, 21, 21, 20, 20, 20, 19, 19, 18, 18, 18, 17, 17, 17,

[0669] 17, 16, 16, 16, 16, 15, 15, 15, 15, 14, 14, 14, 14, 14, 14, 13,

[0670] 13, 13, 13, 13, 12, 12, 12, 12, 12, 12, 12, 11, 11, 11, 11,

[0671] 11, 11, 11, 11, 10, 10, 10, 10, 10, 10, 10,

[0672] Includes the pre-calculated value of 2ALFBitPrecision / NumPt for all possible NumPt - voxel counts in the tab.

[0673] ALFMult[ALFTapsNumPt[tapsidx]] = (1 << ALFBitPrecision) / ALFTapsNumPt[tapsidx]

[0674] The filter application process forms a FilteredPointAttr array and is defined as follows.

[0675] rightshift = ALFBitPrecision + log2(ALFQstep)

[0676] for (tapsidx = 0; tapsidx < NumALFTaps - 1; tapsidx++){

[0677] delta = ALFTapsSumAttributes[tapsidx] - PointAttr[curPtIdx][c]*ALFTapsNumPt[tapsidx]

[0678] term_scaled = ALFAttrCoeff[cubeidx][c][tapsidx]*delta*ALFMult[ALFTapsNumPt[tapsidx]]

[0679] term = (term_scaled*Sign(term_scaled) + (1 << (rightshift - 1)) >> rightshift

[0680] term[tapsidx] = term*Sign(term_scaled)

[0681] FilteredPointAttr[curPtIdx][c] = PointAttr[curPtIdx][c] + term[0] + term[1] + term[2] + term[3]

[0682] FIG. 38 shows an encoding device according to embodiments.

[0683] The device of FIG. 38 is an encoding device according to embodiments and can perform the transmission device (10000) of FIG. 1, a point cloud video encoder (10002), a transmitter (10003), the acquisition-encoding-transmission (20000-20001-20002) of FIG. 2, encoding based on FIG. 11 to 29, syntax information generation of FIG. 30 to 32, filtering of FIG. 33 to 35, syntax information generation of FIG. 36a and FIG. 36b to 37, encoder of FIG. 38, encoding method of FIG. 40, etc.

[0684] Each component constituting the block diagram of the PCC data encoder of FIG. 38 may correspond to hardware, software, a processor, and / or a combination thereof. The input of the decoder includes an encoded geometric information bitstream and an attribute information bitstream, and these bitstreams are decoded to output the restored PCC data.

[0685] Data Input Section: Receives geometry, attributes, and parameters. Coordinate Transformation: Converts the coordinates of the input point cloud into a form suitable for processing. Geometry Information Conversion. Quantization Processing: Quantizes 3D coordinate information to improve compression efficiency.

[0686] Geometry Encoding Related: Voxelization Processing: Converts the point cloud into voxels (3D pixels). Frame Classification: Processes by separating into P-frames and I-frames. Geometry Information Intra Prediction: Predicts geometry information within the current frame. Geometry Information Inter Prediction: Predicts by comparing with the previous frame. Final Frame Generation: Generates the final frame based on the predicted information.

[0687] Attribute Encoding Related: Color Transform Processing: Converts RGB color information into a form suitable for compression. Lossy-Geom Determination: Determines whether to perform lossy compression. For example, if Yes, recoloring is performed. If No, the next step is performed. Attribute Inter-prediction: Performs inter-prediction of attribute information. Attribute Intra Prediction: Performs intra-prediction of attribute information. Apply Adaptive Loop Filter (ALF): Performs filtering to reduce compression artifacts.

[0688] Reconstructed Geometry: Restores geometry information (geometry data). Entropy Encoding: Encodes the final compressed data using entropy coding. As the encoding output, a geometry bitstream and an attribute bitstream are output.

[0689] FIG. 39 shows a decoding device according to embodiments.

[0690] The device of FIG. 39 is a decoding device according to embodiments and can perform the receiving device (10004) of FIG. 1, a receiver (10005), a point cloud video decoder (10006), the transmission-decoding-rendering (20002-20003-20004) of FIG. 2, the decoder of FIG. 7, the receiving device of FIG. 9, the device of FIG. 10, encoding based on FIG. 11 to 29, acquisition of syntax information of FIG. 30 to 32, filtering of FIG. 33 to 35, acquisition of syntax information of FIG. 36a and FIG. 36b to 37, the decoder of FIG. 39, the decoding method of FIG. 41, etc.

[0691] Each component constituting the block diagram of the PCC data decoder of the 3936 may correspond to hardware, software, a processor, and / or a combination thereof. The input of the decoder includes an encoded geometric information bitstream and an attribute information bitstream, and these bitstreams are decoded to output restored PCC data.

[0692] The decoder in Fig. 39 can follow the reverse process of the operation of the decoder in Fig. 38.

[0693] The decoder receives compressed geometry information (Geometry Bitstreams) and compressed attribute information (Attribute Bitstreams).

[0694] Regarding geometry decoding, geometry information entropy encoding: Decodes the compressed geometry bitstream. Determines whether to perform intra-coding. For example, if Yes, perform geometry information intra-predictive restoration; if No, perform geometry information inter-predictive restoration. Reference frame generation: Generates a reference frame using the restored geometry information. Geometry information transformation inverse quantization processing: Restores the quantized data to the original. Inverse coordinate transformation: Restores the transformed coordinates to the original coordinate system.

[0695] Regarding attribute decoding, attribute information Entropy encoding: Decodes the compressed attribute bitstream. Determines whether to intercode. For example, if Yes, perform Attribute Inter Prediction, and if No, proceed with Attribute Intra Prediction.

[0696] Determines whether to use the adaptive loop filter. For example, if Yes, compression artifacts are removed in the ALF application section, and if No, the process proceeds directly to the next step.

[0697] Through decoding, the restored attribute information is output. The original color is restored by performing the reverse process of color conversion.

[0698] As the final decoding output, the restored geometry information and restored attribute information are output.

[0699] FIG. 40 illustrates a encoding method according to embodiments.

[0700] The method according to the embodiments may include the step of encoding geometry data of point cloud data (S4000) and / or the step of encoding attribute data of point cloud data (S4010), etc.

[0701] The step of encoding attribute data (S4010) includes: a step of adaptively loop filtering attribute data, and the step of adaptively loop filtering may further include: a step of searching for neighbors for a voxel containing point cloud data.

[0702] Neighbors are searched based on previously searched voxels within the voxels in the Molton order.

[0703] Referring to the acceleration of neighbor voxel search according to the embodiments, neighbors are searched based on previously searched voxels within the voxels in the molton order.

[0704] Referring to the neighbor search and diagonal neighbor connection according to the embodiments, neighbors include at least one of straight neighbors adjacent to the current voxel or diagonal neighbors connected but not adjacent to the current voxel, straight neighbors are identified based on succeeding borders adjacent to the current voxel, and diagonal neighbors can be identified based on succeeding borders of neighbor voxels for the current voxel.

[0705] Referring to a neighbor skip-based search according to embodiments, the searching step further includes the step of skipping some of the neighbors, and diagonal neighbors may be skipped, or the boundary regarding the largest block among the voxels may be skipped.

[0706] Referring to the method for determining the window size of an ALF according to the embodiments, neighbors are searched based on the window, and the window size can be generated based on the density of the voxels.

[0707] Referring to 'attr_adaptive_loop_filter_enabled', 'attr_alf_component_type', 'attr_alf_neighbor_search_type', and 'attr_alf_window_size_present_flag' in the HLS of FIG. 30, the bitstream includes a set of attribute parameters regarding attribute data, and the set of attribute parameters includes at least one of: a flag indicating whether a loop filter is applied to the attribute data; or information indicating the type of component of the attribute data to which the loop filter is applied; information indicating the type of neighbor voxel search regarding the loop filter; or a flag indicating whether information indicating the window size regarding the loop filter is included in the header of the attribute data unit within the bitstream, and the type of neighbor voxel search may include at least one of all neighbor voxel search or diagonal neighbor voxel skip.

[0708] Referring to 'attr_alf_window_size', 'attr_alf_qstep_log2', and 'attr_alf_geom_shift' in the ADUH of FIG. 32, the bitstream includes an attribute data unit regarding attribute data, the attribute data unit includes an attribute data unit header, and the attribute data unit header may include at least one of: information indicating the size of a window regarding a loop filter; information indicating a quantization step regarding a loop filter; or information indicating a geometry shift regarding a loop filter.

[0709] The encoding method of FIG. 40 is performed by an encoding device. The encoding device includes a memory; and at least one processor connected to the memory; and the at least one processor may be configured to: encode geometry data of point cloud data; and encode attribute data of point cloud data.

[0710] The embodiments further include a computer-readable storage medium for storing a bitstream generated by the method according to FIG. 40.

[0711] The embodiments further include a method comprising the steps of: acquiring a bitstream for point cloud data; generating the bitstream based on the steps of encoding geometry data of the point cloud data and encoding attribute data of the point cloud data; and transmitting data including the bitstream.

[0712] FIG. 41 illustrates a decoding method according to embodiments.

[0713] The method according to the embodiments may include the step of decoding geometry data of point cloud data within a bitstream (S4100) and / or the step of decoding attribute data of point cloud data (S4110), etc.

[0714] The method in Fig. 40 and the method in Fig. 41 can correspond to each other as inverse processes.

[0715] Referring to FIG. 39, attribute decoding, ALF, neighbor search, etc., the step of decoding attribute data (S4110) includes: a step of adaptively loop filtering attribute data, and the step of adaptively loop filtering may further include: a step of searching neighbors for a voxel containing point cloud data.

[0716] Referring to the acceleration of neighbor voxel search according to the embodiments, neighbors are searched based on previously searched voxels within the voxels in the molton order.

[0717] Referring to the neighbor search and diagonal neighbor connection according to the embodiments, neighbors include at least one of straight neighbors adjacent to the current voxel or diagonal neighbors connected but not adjacent to the current voxel, straight neighbors are identified based on succeeding borders adjacent to the current voxel, and diagonal neighbors can be identified based on succeeding borders of neighbor voxels for the current voxel.

[0718] Referring to a neighbor skip-based search according to embodiments, the searching step further includes the step of skipping some of the neighbors, and diagonal neighbors may be skipped, or the boundary regarding the largest block among the voxels may be skipped.

[0719] Referring to the method for determining the window size of an ALF according to the embodiments, neighbors are searched based on the window, and the window size can be generated based on the density of the voxels.

[0720] Referring to 'attr_adaptive_loop_filter_enabled', 'attr_alf_component_type', 'attr_alf_neighbor_search_type', and 'attr_alf_window_size_present_flag' in the HLS of FIG. 30, the bitstream includes a set of attribute parameters regarding attribute data, and the set of attribute parameters includes at least one of: a flag indicating whether a loop filter is applied to the attribute data; or information indicating the type of component of the attribute data to which the loop filter is applied; information indicating the type of neighbor voxel search regarding the loop filter; or a flag indicating whether information indicating the window size regarding the loop filter is included in the header of the attribute data unit within the bitstream, and the type of neighbor voxel search may include at least one of all neighbor voxel search or diagonal neighbor voxel skip.

[0721] Referring to 'attr_alf_window_size', 'attr_alf_qstep_log2', and 'attr_alf_geom_shift' in the ADUH of FIG. 32, the bitstream includes an attribute data unit regarding attribute data, the attribute data unit includes an attribute data unit header, and the attribute data unit header may include at least one of: information indicating the size of a window regarding a loop filter; information indicating a quantization step regarding a loop filter; or information indicating a geometry shift regarding a loop filter.

[0722] The decoding method of FIG. 41 can be performed by a decoding device. The decoding device includes a memory; and at least one processor connected to the memory; and the at least one processor may be configured to: decode geometry data of point cloud data within a bitstream; and decode attribute data of point cloud data.

[0723] The method and apparatus according to the embodiments provide the following technical effects.

[0724] Through the present invention, the visual quality of a reconstructed image or video is significantly improved by applying adaptive filtering. Filtering parameters are adjusted based on local characteristics of the content, such as texture, edges, noise levels, and geometry, thereby enhancing the artifact removal effect more effectively. Bitstream size can be reduced. ALF improves the efficiency of inter-frame prediction by enhancing the quality of the reference frame. This reduces the number of bits required to encode the difference between frames, thereby improving compression efficiency. ALF can be selectively applied to various regions of an image or video frame. This allows for intensive enhancement to be applied to the most necessary parts without unnecessarily increasing computational complexity or bitrate. Both encoders and decoders must implement ALF in a standardized manner to ensure compatibility and consistent quality across various devices and platforms. ALF can be extended to perform cross-component filtering. By enhancing the quality of the reconstructed point cloud, ALF helps reduce the overall bitrate required for transmission or storage, making it useful for bandwidth-constrained applications.

[0725] The embodiments have been described in terms of methods and / or devices, and the description of the methods and the description of the devices may be applied complementarily.

[0726] Although the drawings have been described separately for the convenience of explanation, it is also possible to design a new embodiment by combining the embodiments described in each drawing. Furthermore, designing a computer-readable recording medium containing a program for executing the previously described embodiments, as required by a person skilled in the art, falls within the scope of the embodiments. The apparatus and method according to the embodiments are not limited to the configuration and method of the embodiments described above; rather, the embodiments may be configured by selectively combining all or part of each embodiment to allow for various modifications. Although preferred embodiments have been illustrated and described, the embodiments are not limited to the specific embodiments described above. It is not only possible for a person skilled in the art to make various modifications without departing from the essence of the embodiments claimed in the claims, but such modifications should not be understood individually from the technical concept or perspective of the embodiments.

[0727] Various components of the device of the embodiments may be implemented by hardware, software, firmware, or a combination thereof. Various components of the embodiments may be implemented as a single chip, for example, a single hardware circuit. Depending on the embodiments, the components according to the embodiments may each be implemented as separate chips. Depending on the embodiments, at least one of the components of the device according to the embodiments may be composed of one or more processors capable of executing one or more programs, and one or more programs may include instructions for performing or executing any one or more of the operations / methods according to the embodiments. Executable instructions for performing the methods / operations of the device according to the embodiments may be stored in non-transient CRMs or other computer program products configured to be executed by one or more processors, or may be stored in transient CRMs or other computer program products configured to be executed by one or more processors. Additionally, memory according to the embodiments may be used as a concept that includes not only volatile memory (e.g., RAM, etc.) but also non-volatile memory, flash memory, PROM, etc. In addition, it may also include implementation in the form of carrier waves, such as transmission over the Internet. Furthermore, processor-readable recording media are distributed across networked computer systems, allowing processor-readable code to be stored and executed in a distributed manner.

[0728] In this document, " / " and "," are interpreted as "and / or." For example, "A / B" is interpreted as "A and / or B," and "A, B" is interpreted as "A and / or B." Additionally, "A / B / C" means "at least one of A, B and / or C." Also, "A, B, C" means "at least one of A, B and / or C." Additionally, in this document, "or" is interpreted as "and / or." For example, "A or B" may mean 1) "A" only, 2) "B" only, or 3) "A and B." In other words, "or" in this document may mean "additionally or alternatively."

[0729] Terms such as "first," "second," etc., may be used to describe various components of the embodiments. However, the interpretation of the various components according to the embodiments should not be limited by these terms. These terms are merely used to distinguish one component from another. For example, the first user input signal may be referred to as the second user input signal. Similarly, the second user input signal may be referred to as the first user input signal. The use of these terms should be interpreted as not departing from the scope of the various embodiments. Although the first user input signal and the second user input signal are both user input signals, they do not imply the same user input signals unless clearly indicated in the context.

[0730] The terms used to describe the embodiments are intended for the purpose of describing specific embodiments and are not intended to limit the embodiments. As used in the description of the embodiments and in the claims, the singular is intended to include the plural unless explicitly indicated in the context. Expressions of and / or are used to mean including all possible combinations between the terms. Expressions of include describe the presence of features, numbers, steps, elements, and / or components and do not imply the exclusion of additional features, numbers, steps, elements, and / or components. Conditional expressions such as "if" or "when" used to describe the embodiments are not limited to being optional. It is intended to be interpreted as "when a specific condition is satisfied," "when a related action is performed in response to a specific condition," or "when a related definition is interpreted."

[0731] Additionally, operations according to the embodiments described herein may be performed by a transmitting and receiving device including memory and / or a processor, depending on the embodiments. The memory may store programs for processing / controlling operations according to the embodiments, and the processor may control various operations described in this document. The processor may be referred to as a controller, etc. Operations in the embodiments may be performed by firmware, software, and / or a combination thereof, and the firmware, software, and / or a combination thereof may be stored in the processor or in memory.

[0732] Meanwhile, the operation according to the embodiments described above may be performed by a transmitting device and / or a receiving device according to the embodiments. The transmitting and receiving device may include a transmitting and receiving unit for transmitting and receiving media data, a memory for storing instructions (program code, algorithm, flowchart and / or data) for a process according to the embodiments, and a processor for controlling the operations of the transmitting and receiving devices.

[0733] The processor may be referred to as a controller, etc., and may correspond, for example, to hardware, software, and / or a combination thereof. The operation according to the embodiments described above may be performed by the processor. Additionally, the processor may be implemented as an encoder / decoder, etc., for the operation of the embodiments described above.

[0734] As described above, the relevant details have been explained in the best mode for carrying out the embodiments.

[0735] As described above, the embodiments may be applied wholly or partially to point cloud data transmission and reception devices and systems.

[0736] Those skilled in the art may make various changes or modifications to the embodiments within the scope of the embodiments.

[0737] The embodiments may include modifications / variations, and such modifications / variations do not exceed the scope of the claims and their equivalents.< / int> < / int>

Claims

1. A step of decoding geometry data of point cloud data within a bitstream; and A step of decoding attribute data of the above point cloud data; comprising Decryption method.

2. In Paragraph 1, The step of decoding the above attribute data is: The method includes the step of adaptively loop filtering the above attribute data, The above adaptive loop filtering step is: The method further comprises the step of searching for neighbors for a voxel containing the above point cloud data. Decryption method.

3. In Paragraph 2, The above neighbors are searched based on previously searched voxels within the voxels in the Molton order, Decryption method.

4. In Paragraph 3, The above neighbors include at least one of a straight neighbor adjacent to the current voxel or a diagonal neighbor that is not adjacent to the current voxel but is connected, The above linear neighbors are identified based on the succeeding border adjacent to the current voxel, and The above diagonal neighbors are identified based on the continuous boundary of neighbor voxels for the above current voxel, Decryption method.

5. In Paragraph 4, The step of searching for the above neighbors is: It further includes a step of skipping some of the above neighbors, and The above diagonal neighbor is skipped, or the boundary regarding the largest block among the above voxels is skipped, Decryption method.

6. In Paragraph 5, The above neighbors are searched based on the window, and The size of the above window is generated based on the voxel density, Decryption method.

7. In Paragraph 1, The above bitstream includes a set of attribute parameters regarding the attribute data, and The above set of attribute parameters is: A flag indicating whether a loop filter is applied to the above attribute data; or Information indicating the type of the component of the attribute data to which the above loop filter is applied; Information indicating the type of neighbor voxel search regarding the above loop filter; or It includes at least one of a flag indicating whether information indicating the window size regarding the loop filter is included in the header of the attribute data unit within the bitstream, and The type of the above neighbor voxel search includes at least one of all neighbor voxel searches or diagonal neighbor voxel skips, Decryption method.

8. In Paragraph 1, The above bitstream includes an attribute data unit regarding the attribute data, and The above attribute data unit includes an attribute data unit header, and The above attribute data unit header is: Information indicating the size of the window regarding the loop filter; Information indicating a quantization step regarding the above-mentioned loop filter; or Information indicating a geometry shift regarding the loop filter above; comprising at least one of Decryption method.

9. Memory; and At least one processor connected to the memory; comprising, wherein the at least one processor: Decoding geometry data of point cloud data within a bitstream; and Decoding attribute data of the above point cloud data; configured to do so, Decoding device.

10. A step of encoding the geometry data of the point cloud data; and A step of encoding attribute data of the above point cloud data; comprising Encoding method.

11. In Paragraph 10, The step of encoding the above attribute data is: The method includes the step of adaptively loop filtering the above attribute data, The above adaptive loop filtering step is: The method further comprises the step of searching for neighbors for a voxel containing the above point cloud data. Encoding method.

12. In Paragraph 11, The above neighbors are searched based on previously searched voxels within the voxels in the Molton order, Encoding method.

13. Memory; and At least one processor connected to the memory; comprising, wherein the at least one processor: Encoding the geometry data of the point cloud data; and Configured to encode the attribute data of the above point cloud data; Decoding device.

14. A computer-readable storage medium for storing a bitstream generated by the method according to paragraph 10.

15. Step for acquiring a bitstream for point cloud data, The bitstream is generated based on the step of encoding geometry data of the point cloud data; and the step of encoding attribute data of the point cloud data; and A method comprising the step of transmitting data including the bitstream above.

Images