Method, apparatus, and medium for point cloud processing
The method improves point cloud coding efficiency by using low frequency information to perform adaptive downsampling and upsampling, addressing the limitations of existing techniques in maintaining reconstruction quality during compression.
Patent Information
- Application Number
- PCT/CN2024/118614
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-09-13
- Filing Date
- 2024-09-12
- Publication Date
- 2025-03-20
AI Technical Summary
Existing point cloud coding techniques face challenges in improving coding efficiency, particularly in effectively downsampling and upsampling point clouds while maintaining reconstruction quality.
The proposed method involves obtaining a first set of points representing low frequency information from a current point cloud sample, and using these points to perform collaborative adaptive downsampling and upsampling, thereby improving coding efficiency.
This approach enhances coding efficiency by utilizing collaborative adaptive downsampling and upsampling, allowing for better compression of point clouds while maintaining reconstruction quality.
Smart Images

Figure CN2024118614_20032025_PF_FP_ABST
Abstract
Description
METHOD, APPARATUS, AND MEDIUM FOR POINT CLOUD PROCESSING
[0001] FIELDS
[0002] Embodiments of the present disclosure relates generally to point cloud processing techniques, and more particularly, to collaborative adaptive downsampling and upsampling of point clouds.BACKGROUND
[0003] A point cloud is a collection of individual data points in a three-dimensional (3D) plane with each point having a set coordinate on the X, Y, and Z axes. Thus, a point cloud may be used to represent the physical content of the three-dimensional space. Point clouds have shown to be a promising way to represent 3D visual data for a wide range of immersive applications, from augmented reality to autonomous cars.
[0004] Point cloud coding standards have evolved primarily through the development of the well-known MPEG organization. MPEG, short for Moving Picture Experts Group, is one of the main standardization groups dealing with multimedia. In 2017, the MPEG 3D Graphics Coding group (3DG) published a call for proposals (CFP) document to start to develop point cloud coding standard. The final standard will consist in two classes of solutions. Video-based Point Cloud Compression (V-PCC or VPCC) is appropriate for point sets with a relatively uniform distribution of points. Geometry-based Point Cloud Compression (G-PCC or GPCC) is appropriate for more sparse distributions. However, coding efficiency of conventional point cloud coding techniques is generally expected to be further improved.SUMMARY
[0005] Embodiments of the present disclosure provide a solution for point cloud processing.
[0006] In a first aspect, a method for point cloud processing is proposed. The method comprises: obtaining, for a conversion between a current point cloud (PC) sample of a point cloud sequence and a bitstream of the point cloud sequence, a first set of points in the current PC sample, the first set of points representing low frequency information of the current PC sample; and performing the conversion based on the first set of points.
[0007] Based on the method in accordance with the first aspect of the present disclosure, a point cloud is coded based on a set of points representing low frequency information of the point cloud. Compared with the conventional solution, the proposed method can advantageously utilize collaborative adaptive downsampling and upsampling of a point cloud to assist in coding point cloud. Thereby, the coding efficiency can be improved.
[0008] In a second aspect, an apparatus for point cloud processing is proposed. The apparatus comprises a processor and a non-transitory memory with instructions thereon. The instructions upon execution by the processor, cause the processor to perform a method in accordance with the first aspect of the present disclosure.
[0009] In a third aspect, a non-transitory computer-readable storage medium is proposed. The non-transitory computer-readable storage medium stores instructions that cause a processor to perform a method in accordance with the first aspect of the present disclosure.
[0010] In a fourth aspect, another non-transitory computer-readable recording medium is proposed. The non-transitory computer-readable recording medium stores a bitstream of a point cloud sequence which is generated by a method performed by an apparatus for point cloud processing. The method comprises: obtaining a first set of points in a current PC sample of the point cloud sequence, the first set of points representing low frequency information of the current PC sample; and generating the bitstream based on the first set of points.
[0011] In a fifth aspect, a method for storing a bitstream of a point cloud sequence is proposed. The method comprises: obtaining a first set of points in a current PC sample of the point cloud sequence, the first set of points representing low frequency information of the current PC sample; generating the bitstream based on the first set of points; and storing the bitstream in a non-transitory computer-readable recording medium.
[0012] This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Through the following detailed description with reference to the accompanying drawings, the above and other objectives, features, and advantages of example embodiments of the present disclosure will become more apparent. In the example embodiments of the present disclosure, the same reference numerals usually refer to the same components.
[0014] Fig. 1 is a block diagram that illustrates an example point cloud coding system that may utilize the techniques of the present disclosure;
[0015] Fig. 2 illustrates a block diagram that illustrates an example point cloud encoder, in accordance with some embodiments of the present disclosure;
[0016] Fig. 3 illustrates a block diagram that illustrates an example point cloud decoder, in accordance with some embodiments of the present disclosure;
[0017] Fig. 4 illustrates an example flow of the collaborative adaptive downsampling and upsampling method of point clouds in accordance with embodiments of the present disclosure;
[0018] Fig. 5 illustrates a flowchart of a method for point cloud processing in accordance with embodiments of the present disclosure; and
[0019] Fig. 6 illustrates a block diagram of a computing device in which various embodiments of the present disclosure can be implemented.
[0020] Throughout the drawings, the same or similar reference numerals usually refer to the same or similar elements.DETAILED DESCRIPTION
[0021] Principle of the present disclosure will now be described with reference to some embodiments. It is to be understood that these embodiments are described only for the purpose of illustration and help those skilled in the art to understand and implement the present disclosure, without suggesting any limitation as to the scope of the disclosure. The disclosure described herein can be implemented in various manners other than the ones described below.
[0022] In the following description and claims, unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skills in the art to which this disclosure belongs.
[0023] References in the present disclosure to “one embodiment, ” “an embodiment, ” “an example embodiment, ” and the like indicate that the embodiment described may include a particular feature, structure, or characteristic, but it is not necessary that every embodiment includes the particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same embodiment. Further, when a particular feature, structure, or characteristic is described in connection with an example embodiment, it is submitted that it is within the knowledge of one skilled in the art to affect such feature, structure, or characteristic in connection with other embodiments whether or not explicitly described.
[0024] It shall be understood that although the terms “first” and “second” etc. may be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, a first element could be termed a second element, and similarly, a second element could be termed a first element, without departing from the scope of example embodiments. As used herein, the term “and / or” includes any and all combinations of one or more of the listed terms.
[0025] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments. As used herein, the singular forms “a” , “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprises” , “comprising” , “has” , “having” , “includes” and / or “including” , when used herein, specify the presence of stated features, elements, and / or components etc., but do not preclude the presence or addition of one or more other features, elements, components and / or combinations thereof.
[0026] Example Environment
[0027] Fig. 1 is a block diagram that illustrates an example point cloud coding system 100 that may utilize the techniques of the present disclosure. As shown, the point cloud coding system 100 may include a source device 110 and a destination device 120. The source device 110 can be also referred to as a point cloud encoding device, and the destination device 120 can be also referred to as a point cloud decoding device. In operation, the source device 110 can be configured to generate encoded point cloud data and the destination device 120 can be configured to decode the encoded point cloud data generated by the source device 110. The techniques of this disclosure are generally directed to coding (encoding and / or decoding) point cloud data, i.e., to support point cloud compression. The coding may be effective in compressing and / or decompressing point cloud data.
[0028] Source device 100 and destination device 120 may comprise any of a wide range of devices, including desktop computers, notebook (i.e., laptop) computers, tablet computers, set-top boxes, telephone handsets such as smartphones and mobile phones, televisions, cameras, display devices, digital media players, video gaming consoles, video streaming devices, vehicles (e.g., terrestrial or marine vehicles, spacecraft, aircraft, etc. ) , robots, LIDAR devices, satellites, extended reality devices, or the like. In some cases, source device 100 and destination device 120 may be equipped for wireless communication.
[0029] The source device 100 may include a data source 112, a memory 114, a GPCC encoder 116, and an input / output (I / O) interface 118. The destination device 120 may include an input / output (I / O) interface 128, a GPCC decoder 126, a memory 124, and a data consumer 122. In accordance with this disclosure, GPCC encoder 116 of source device 100 and GPCC decoder 126 of destination device 120 may be configured to apply the techniques of this disclosure related to point cloud coding. Thus, source device 100 represents an example of an encoding device, while destination device 120 represents an example of a decoding device. In other examples, source device 100 and destination device 120 may include other components or arrangements. For example, source device 100 may receive data (e.g., point cloud data) from an internal or external source. Likewise, destination device 120 may interface with an external data consumer, rather than include a data consumer in the same device.
[0030] In general, data source 112 represents a source of point cloud data (i.e., raw, unencoded point cloud data) and may provide a sequential series of “frames” of the point cloud data to GPCC encoder 116, which encodes point cloud data for the frames. In some examples, data source 112 generates the point cloud data. Data source 112 of source device 100 may include a point cloud capture device, such as any of a variety of cameras or sensors, e.g., one or more video cameras, an archive containing previously captured point cloud data, a 3D scanner or a light detection and ranging (LIDAR) device, and / or a data feed interface to receive point cloud data from a data content provider. Thus, in some examples, data source 112 may generate the point cloud data based on signals from a LIDAR apparatus. Alternatively or additionally, point cloud data may be computer-generated from scanner, camera, sensor or other data. For example, data source 112 may generate the point cloud data, or produce a combination of live point cloud data, archived point cloud data, and computer-generated point cloud data. In each case, GPCC encoder 116 encodes the captured, pre-captured, or computer-generated point cloud data. GPCC encoder 116 may rearrange frames of the point cloud data from the received order (sometimes referred to as “display order” ) into a coding order for coding. GPCC encoder 116 may generate one or more bitstreams including encoded point cloud data. Source device 100 may then output the encoded point cloud data via I / O interface 118 for reception and / or retrieval by, e.g., I / O interface 128 of destination device 120. The encoded point cloud data may be transmitted directly to destination device 120 via the I / O interface 118 through the network 130A. The encoded point cloud data may also be stored onto a storage medium / server 130B for access by destination device 120.
[0031] Memory 114 of source device 100 and memory 124 of destination device 120 may represent general purpose memories. In some examples, memory 114 and memory 124 may store raw point cloud data, e.g., raw point cloud data from data source 112 and raw, decoded point cloud data from GPCC decoder 126. Additionally or alternatively, memory 114 and memory 124 may store software instructions executable by, e.g., GPCC encoder 116 and GPCC decoder 126, respectively. Although memory 114 and memory 124 are shown separately from GPCC encoder 116 and GPCC decoder 126 in this example, it should be understood that GPCC encoder 116 and GPCC decoder 126 may also include internal memories for functionally similar or equivalent purposes. Furthermore, memory 114 and memory 124 may store encoded point cloud data, e.g., output from GPCC encoder 116 and input to GPCC decoder 126. In some examples, portions of memory 114 and memory 124 may be allocated as one or more buffers, e.g., to store raw, decoded, and / or encoded point cloud data. For instance, memory 114 and memory 124 may store point cloud data.
[0032] I / O interface 118 and I / O interface 128 may represent wireless transmitters / receivers, modems, wired networking components (e.g., Ethernet cards) , wireless communication components that operate according to any of a variety of IEEE 802.11 standards, or other physical components. In examples where I / O interface 118 and I / O interface 128 comprise wireless components, I / O interface 118 and I / O interface 128 may be configured to transfer data, such as encoded point cloud data, according to a cellular communication standard, such as 4G, 4G-LTE (Long-Term Evolution) , LTE Advanced, 5G, or the like. In some examples where I / O interface 118 comprises a wireless transmitter, I / O interface 118 and I / O interface 128 may be configured to transfer data, such as encoded point cloud data, according to other wireless standards, such as an IEEE 802.11 specification. In some examples, source device 100 and / or destination device 120 may include respective system-on-a-chip (SoC) devices. For example, source device 100 may include an SoC device to perform the functionality attributed to GPCC encoder 116 and / or I / O interface 118, and destination device 120 may include an SoC device to perform the functionality attributed to GPCC decoder 126 and / or I / O interface 128.
[0033] The techniques of this disclosure may be applied to encoding and decoding in support of any of a variety of applications, such as communication between autonomous vehicles, communication between scanners, cameras, sensors and processing devices such as local or remote servers, geographic mapping, or other applications.
[0034] I / O interface 128 of destination device 120 receives an encoded bitstream from source device 110. The encoded bitstream may include signaling information defined by GPCC encoder 116, which is also used by GPCC decoder 126, such as syntax elements having values that represent a point cloud. Data consumer 122 uses the decoded data. For example, data consumer 122 may use the decoded point cloud data to determine the locations of physical objects. In some examples, data consumer 122 may comprise a display to present imagery based on the point cloud data.
[0035] GPCC encoder 116 and GPCC decoder 126 each may be implemented as any of a variety of suitable encoder and / or decoder circuitry, such as one or more microprocessors, digital signal processors (DSPs) , application specific integrated circuits (ASICs) , field programmable gate arrays (FPGAs) , discrete logic, software, hardware, firmware or any combinations thereof. When the techniques are implemented partially in software, a device may store instructions for the software in a suitable, non-transitory computer-readable medium and execute the instructions in hardware using one or more processors to perform the techniques of this disclosure. Each of GPCC encoder 116 and GPCC decoder 126 may be included in one or more encoders or decoders, either of which may be integrated as part of a combined encoder / decoder (CODEC) in a respective device. A device including GPCC encoder 116 and / or GPCC decoder 126 may comprise one or more integrated circuits, microprocessors, and / or other types of devices.
[0036] GPCC encoder 116 and GPCC decoder 126 may operate according to a coding standard, such as video point cloud compression (VPCC) standard or a geometry point cloud compression (GPCC) standard. This disclosure may generally refer to coding (e.g., encoding and decoding) of frames to include the process of encoding or decoding data. An encoded bitstream generally includes a series of values for syntax elements representative of coding decisions (e.g., coding modes) .
[0037] A point cloud may contain a set of points in a 3D space, and may have attributes associated with the point. The attributes may be color information such as R, G, B or Y, Cb, Cr, or reflectance information, or other attributes. Point clouds may be captured by a variety of cameras or sensors such as LIDAR sensors and 3D scanners and may also be computer-generated. Point cloud data are used in a variety of applications including, but not limited to, construction (modeling) , graphics (3D models for visualizing and animation) , and the automotive industry (LIDAR sensors used to help in navigation) .
[0038] Fig. 2 is a block diagram illustrating an example of a GPCC encoder 200, which may be an example of the GPCC encoder 116 in the system 100 illustrated in Fig. 1, in accordance with some embodiments of the present disclosure. Fig. 3 is a block diagram illustrating an example of a GPCC decoder 300, which may be an example of the GPCC decoder 126 in the system 100 illustrated in Fig. 1, in accordance with some embodiments of the present disclosure.
[0039] In both GPCC encoder 200 and GPCC decoder 300, point cloud positions are coded first. Attribute coding depends on the decoded geometry. In Fig. 2 and Fig. 3, the region adaptive hierarchical transform (RAHT) unit 218, surface approximation analysis unit 212, RAHT unit 314 and surface approximation synthesis unit 310 are options typically used for Category 1 data. The level-of-detail (LOD) generation unit 220, lifting unit 222, LOD generation unit 316 and inverse lifting unit 318 are options typically used for Category 3 data. All the other units are common between Categories 1 and 3.
[0040] For Category 3 data, the compressed geometry is typically represented as an octree from the root all the way down to a leaf level of individual voxels. For Category 1 data, the compressed geometry is typically represented by a pruned octree (i.e., an octree from the root down to a leaf level of blocks larger than voxels) plus a model that approximates the surface within each leaf of the pruned octree. In this way, both Category 1 and 3 data share the octree coding mechanism, while Category 1 data may in addition approximate the voxels within each leaf with a surface model. The surface model used is a triangulation comprising 1-10 triangles per block, resulting in a triangle soup. The Category 1 geometry codec is therefore known as the Trisoup geometry codec, while the Category 3 geometry codec is known as the Octree geometry codec.
[0041] In the example of Fig. 2, GPCC encoder 200 may include a coordinate transform unit 202, a color transform unit 204, a voxelization unit 206, an attribute transfer unit 208, an octree analysis unit 210, a surface approximation analysis unit 212, an arithmetic encoding unit 214, a geometry reconstruction unit 216, an RAHT unit 218, a LOD generation unit 220, a lifting unit 222, a coefficient quantization unit 224, and an arithmetic encoding unit 226.
[0042] As shown in the example of Fig. 2, GPCC encoder 200 may receive a set of positions and a set of attributes. The positions may include coordinates of points in a point cloud. The attributes may include information about points in the point cloud, such as colors associated with points in the point cloud.
[0043] Coordinate transform unit 202 may apply a transform to the coordinates of the points to transform the coordinates from an initial domain to a transform domain. This disclosure may refer to the transformed coordinates as transform coordinates. Color transform unit 204 may apply a transform to convert color information of the attributes to a different domain. For example, color transform unit 204 may convert color information from an RGB color space to a YCbCr color space.
[0044] Furthermore, in the example of Fig. 2, voxelization unit 206 may voxelize the transform coordinates. Voxelization of the transform coordinates may include quantizing and removing some points of the point cloud. In other words, multiple points of the point cloud may be subsumed within a single “voxel, ” which may thereafter be treated in some respects as one point. Furthermore, octree analysis unit 210 may generate an octree based on the voxelized transform coordinates. Additionally, in the example of Fig. 2, surface approximation analysis unit 212 may analyze the points to potentially determine a surface representation of sets of the points. Arithmetic encoding unit 214 may perform arithmetic encoding on syntax elements representing the information of the octree and / or surfaces determined by surface approximation analysis unit 212. GPCC encoder 200 may output these syntax elements in a geometry bitstream.
[0045] Geometry reconstruction unit 216 may reconstruct transform coordinates of points in the point cloud based on the octree, data indicating the surfaces determined by surface approximation analysis unit 212, and / or other information. The number of transform coordinates reconstructed by geometry reconstruction unit 216 may be different from the original number of points of the point cloud because of voxelization and surface approximation. This disclosure may refer to the resulting points as reconstructed points. Attribute transfer unit 208 may transfer attributes of the original points of the point cloud to reconstructed points of the point cloud data.
[0046] Furthermore, RAHT unit 218 may apply RAHT coding to the attributes of the reconstructed points. Alternatively or additionally, LOD generation unit 220 and lifting unit 222 may apply LOD processing and lifting, respectively, to the attributes of the reconstructed points. RAHT unit 218 and lifting unit 222 may generate coefficients based on the attributes. Coefficient quantization unit 224 may quantize the coefficients generated by RAHT unit 218 or lifting unit 222. Arithmetic encoding unit 226 may apply arithmetic coding to syntax elements representing the quantized coefficients. GPCC encoder 200 may output these syntax elements in an attribute bitstream.
[0047] In the example of Fig. 3, GPCC decoder 300 may include a geometry arithmetic decoding unit 302, an attribute arithmetic decoding unit 304, an octree synthesis unit 306, an inverse quantization unit 308, a surface approximation synthesis unit 310, a geometry reconstruction unit 312, a RAHT unit 314, a LOD generation unit 316, an inverse lifting unit 318, a coordinate inverse transform unit 320, and a color inverse transform unit 322.
[0048] GPCC decoder 300 may obtain a geometry bitstream and an attribute bitstream. Geometry arithmetic decoding unit 302 of decoder 300 may apply arithmetic decoding (e.g., CABAC or other type of arithmetic decoding) to syntax elements in the geometry bitstream. Similarly, attribute arithmetic decoding unit 304 may apply arithmetic decoding to syntax elements in attribute bitstream.
[0049] Octree synthesis unit 306 may synthesize an octree based on syntax elements parsed from geometry bitstream. In instances where surface approximation is used in geometry bitstream, surface approximation synthesis unit 310 may determine a surface model based on syntax elements parsed from geometry bitstream and based on the octree.
[0050] Furthermore, geometry reconstruction unit 312 may perform a reconstruction to determine coordinates of points in a point cloud. Coordinate inverse transform unit 320 may apply an inverse transform to the reconstructed coordinates to convert the reconstructed coordinates (positions) of the points in the point cloud from a transform domain back into an initial domain.
[0051] Additionally, in the example of Fig. 3, inverse quantization unit 308 may inverse quantize attribute values. The attribute values may be based on syntax elements obtained from attribute bitstream (e.g., including syntax elements decoded by attribute arithmetic decoding unit 304) .
[0052] Depending on how the attribute values are encoded, RAHT unit 314 may perform RAHT coding to determine, based on the inverse quantized attribute values, color values for points of the point cloud. Alternatively, LOD generation unit 316 and inverse lifting unit 318 may determine color values for points of the point cloud using a level of detail-based technique.
[0053] Furthermore, in the example of Fig. 3, color inverse transform unit 322 may apply an inverse color transform to the color values. The inverse color transform may be an inverse of a color transform applied by color transform unit 204 of encoder 200. For example, color transform unit 204 may transform color information from an RGB color space to a YCbCr color space. Accordingly, color inverse transform unit 322 may transform color information from the YCbCr color space to the RGB color space.
[0054] The various units of Fig. 2 and Fig. 3 are illustrated to assist with understanding the operations performed by encoder 200 and decoder 300. The units may be implemented as fixed-function circuits, programmable circuits, or a combination thereof. Fixed-function circuits refer to circuits that provide particular functionality and are preset on the operations that can be performed. Programmable circuits refer to circuits that can be programmed to perform various tasks and provide flexible functionality in the operations that can be performed. For instance, programmable circuits may execute software or firmware that cause the programmable circuits to operate in the manner defined by instructions of the software or firmware. Fixed-function circuits may execute software instructions (e.g., to receive parameters or output parameters) , but the types of operations that the fixed-function circuits perform are generally immutable. In some examples, one or more of the units may be distinct circuit blocks (fixed-function or programmable) , and in some examples, one or more of the units may be integrated circuits.
[0055] Some exemplary embodiments of the present disclosure will be described in detailed hereinafter. It should be understood that section headings are used in the present document to facilitate ease of understanding and do not limit the embodiments disclosed in a section to only that section. Furthermore, while certain embodiments are described with reference to GPCC or other specific point cloud codecs, the disclosed techniques are applicable to other point cloud coding technologies also. Furthermore, while some embodiments describe point cloud coding steps in detail, it will be understood that corresponding steps decoding that undo the coding will be implemented by a decoder.
[0056] 1. Brief Summary
[0057] This disclosure is related to point cloud coding pre-processing and post-processing technologies. Specifically, it is related to learning-based point cloud geometry downsampling and upsampling. The ideas may be combined with point cloud coding standard, e.g., the being-developed Geometry based Point Cloud Compression (G-PCC) .
[0058] 2. Abbreviations
[0059] G-PCC Geometry based Point Cloud Compression
[0060] MPEG Moving Picture Experts Group
[0061] 3. Introduction
[0062] In point cloud compression, traditional octree coding, grid coding, mapping coding, and attribute coding have provided the basic ideas and framework for compression. Following their encoding principles and module structures, people use various signal processing methods to design new modules or optimize and enhance the old ones. The same is true for learning-based point cloud compression, which can replace traditional modules using neural network models and also optimize model parameters based on data-driven optimization.
[0063] At present, both the traditional point cloud compression method and the learning-based point cloud compression method try to retain all the information of the original point cloud as much as the code rate allows, but not all the points of the original point cloud are important and deserve to be preserved. Downsampling pre-processing methods can effectively reduce the information redundancy of the original point cloud. The code rate of point cloud can be greatly reduced by down-sampling pre-processing, while the original reconstruction quality can be maintained as much as possible by up-sampling post-processing.
[0064] Farthest point sampling (FPS) has been widely used as a pooling operation in point cloud neural processing systems. However, FPS does not take into account the further processing of the sampled points and may result in sub-optimal performance. Recently, alternative sub-sampling methods have been proposed. An existing design introduced a critical points layer, which passes on points with the most active features to the next network layer. An existing design used Gumbel subset sampling during the training of a classification network instead of FPS, to improve its accuracy. For upsampling methods, most of them are based on pointnet base operations. An existing design introduced PU-Net, which learns multi-scale features per point and expands the point set via multi-branch MLPs. However, PU-Net needs to downsample the input first to learn multi-scale features, which causes unnecessary resolution loss., an existing design proposed PU-GAN, a Generative Adverserial Network (GAN) designed to learn upsampled point distributions. While the major contribution and the performance gain is from the discriminator part, the generator architecture receives less attention in their work. All of the current upsampling and downsampling methods are performed separately, there is no joint upsampling and downsampling method, and this method is more suitable for the compression task, which can keep the original reconstruction quality as much as possible while lowering the bit rate.
[0065] 3.1 Sparse Convolution
[0066] To exploit the sparsity of point clouds, scholars have conducted many explorations, such as octree-based CNNs and sparse CNNs. For sparse CNNs, the data tensor is represented by a set of coordinates C and the associated features F. The convolution aggregates only the features that are positively occupying the coordinates. It is defined as:
[0067] where Cin and Cout are input coordinates and output coordinates. and are input and output feature vectors at coordinate t. N3 (t, Cin) = {i |t+i∈Cin, i∈N3} defines a 3D convolutional kernel, covering a set of locations centered at t with offset i's in Cin. Wi denotes the kernel value at offset i. This sparse convolution exploits the sparsity of the point cloud to reduce the complexity and computes only on the positively occupied voxels.
[0068] 4. Problems
[0069] The existing learning-based point cloud downsampling and upsampling methods have the following problems:
[0070] 1. The current downsampling methods and upsampling methods are split and not well combined to achieve the reconstruction task and used for the post-processing and the pre-processing of the compression.
[0071] 2. Most of the current down sampling methods are point-based methods, and the low-resolution point clouds obtained in this way are not a subset of the original point clouds with errors. Moreover, these methods are relatively expensive and can only handle small-scale point clouds.
[0072] 3. Most of the upsampling methods use features to regress the coordinates of the point cloud directly, which makes the upsampling reconstruction difficult and error-prone.
[0073] 4. The current upsampling and downsampling training mechanism is single, and it is not possible to get a more adapted upsampling network and downsampling network.
[0074] 5. Detailed Solutions
[0075] To solve the above problems and some other problems not mentioned, methods as summarized below are disclosed. The inventions should be considered as examples to explain the general concepts and should not be interpreted in a narrow way. Furthermore, these inventions can be applied individually or combined in any manner.
[0076] In the following discussions, the term “encoder” refers to the model to code of the information to be signalled. The term “decoder” refers to the model to decode the compression bits to get the signalled information.
[0077] 1) It is proposed to down sample the point cloud to obtain low frequency information.
[0078] a. In one example, the low frequency information may be evaluated in terms of the geometric character of point cloud, such as geometric structure or geometric normals.
[0079] i. Furthermore, the low frequency information in the point cloud may be evaluated using the geometric structure.
[0080] 1. In one example, the point set (such as uniform point set, farthest sampled point set, etc. ) may be used to obtain the backbone structure of the point cloud.
[0081] a. In one example, the farthest sampling method may be used to obtain the farthest sampled point set.
[0082] 2. In one example, the voxelized point clouds may be used to characterize the geometric structure of the point cloud.
[0083] a. In one example, sparse point clouds with large granularity voxelization are represented as low-frequency information.
[0084] b. In one example, the low frequency information may be evaluated in terms of the attribute character of point cloud, such as color information.
[0085] c. In one example, the low frequency information may be evaluated in terms of both the geometric character and the attribute character of point cloud.
[0086] d. In one example, the low frequency information may be obtained using a traditional methods or a learning-based approach.
[0087] i. In one example, a neural network-based learning approach may be used to obtain the low frequency information.
[0088] 1. In one example, the sparse convolution may be used as the basic operation in the convolutional network.
[0089] a. In one example, the step size of sparse convolution may be N. For example, N is equal to 2.
[0090] b. In one example, N may be pre-defined.
[0091] c. In one example, N may be siganlled.
[0092] 2) It is proposed to down sample the point cloud to obtain high frequency features of the point cloud.
[0093] a. In one example, the high frequency information in the point cloud may be evaluated using the geometric structure.
[0094] i. In one example, the points lost by the downsampling operation may be represented as high-frequency information.
[0095] b. In one example, high-frequency features can be obtained by hand-designed operators or learning-based approach from high frequency information.
[0096] i. In one example, a neural learning approach may be used to obtain the high frequency features.
[0097] c. In one example, high frequency features may include local high frequency features and global high frequency features.
[0098] i. In one example, neural learning approachs may be used to obtain the local high frequency features and global high frequency features.
[0099] ii. In one example, the local high frequency information may be obtained by graph convolution.
[0100] 1. In one example, the graph convolution can form graphs according to k nearest neighbours.
[0101] a. In one example, the graphs in the graph convolution may be obtained by finding k-nearest neighbours in Euclidean space.
[0102] b. In one example, the graphs in graph convolution may be obtained by finding k-nearest neighbours in feature space.
[0103] 2. In one example, features within a local graph may be aggregated using summation over all points in the graph.
[0104] 3. In one example, features within a local graph may be aggregated by weighting the features of all points within the graph.
[0105] a. In one example, feature weighted aggregation may be performed based on geometric distances.
[0106] b. In one example, feature weighted aggregation may be performed based on feature similarity.
[0107] 4. In one example, the sparse convolution may be used as the basic operation in the graph aggregation.
[0108] a. In one example, the features obtained by convolutional downsampling can be considered as high frequency features.
[0109] b. In one example, the step size of sparse convolution may be N.For example, N is equal to 2.
[0110] c. In one example, N may be pre-defined.
[0111] d. In one example, N may be signalled.
[0112] iii. In one example, the global high frequency information may be obtained by graph aggregation.
[0113] 1. In one example, the global graph of global graph aggregation may be acquired over the overall point cloud.
[0114] a. In one example, the global graph may be obtained by non-learning approach.
[0115] i. In one example, global graph may be obtained by farthest point sampling.
[0116] b. In one example, the global graph can be obtained by a learning-based approach.
[0117] i. In one example, based on the response value of each point obtained by the network, the global graph is formed by the first k high points.
[0118] 1. For example, k is equal to 8.
[0119] 2. In one example, features within the global graph may be aggregated using summation over all points in the graph.
[0120] 3. In one example, features within the graph may be aggregated by weighting the features of all points within the graph.
[0121] a. In one example, feature weighted aggregation may be performed based on geometric distances.
[0122] b. In one example, feature weighted aggregation may be performed based on feature similarity.
[0123] 4. In one example, the sparse convolution may be used as the basic operation in the graph aggregation.
[0124] a. In one example, the features obtained by convolutional downsampling can be considered as high frequency features.
[0125] b. In one example, the step size of sparse convolution may be N. For example, N is equal to 2.
[0126] c. In one example, N may be pre-defined.
[0127] d. In one example, N may be signalled.
[0128] 3) It is proposed to code and signal the final sampled point cloud to the decoder.
[0129] a. In one example, the final sampled point cloud may be coded by point cloud codec.
[0130] i. In one example, the point cloud codec may be G-PCC, V-PCC, Draco etc.
[0131] 4) It is proposed to code and signal the features to the decoder.
[0132] a. In one example, the features may be coded with fixed-length coding, unary coding, truncated unary coding, etc. al.
[0133] b. In one example, the features may be coding in a predictive way.
[0134] 5) It is proposed to predict the high frequency features using initial features.
[0135] a. In one example, the prediction of high frequency features may be using convolution operation to expand the feature dimension only.
[0136] b. In one example, the prediction of high frequency features may be implemented by a general feature extraction network.
[0137] c. In one example, the prediction of high frequency features may be performed by a generative adversarial network which contains a generator and a discriminator.
[0138] i. In on example, the generator may be performed using the convolutional block with feature expansion operations.
[0139] 1. In one example, the convolutional block may be implemented using sparse convolution.
[0140] a. In one example, multiple consecutive sparse convolution operations may be used. For example, three consecutive sparse convolution operations may be used.
[0141] ii. In on example, the discriminator may be performed using the convolutional block as a basic operation.
[0142] 1. In one example, the convolutional block may be implemented using sparse convolution.
[0143] a. In one example, multiple consecutive sparse convolution operations may be used. For example, three consecutive sparse convolution operations may be used.
[0144] 2. In one example, the maximum pooling operation may be utilized to retain the detail information.
[0145] 3. In one example, the average pooling operation may be employed to retain the overall structure information.
[0146] d. In on example, the predicted high frequency features can be constrained by an effective loss function.
[0147] i. In on example, L2 distance loss function may be used to constrain the difference between the predicted high-frequency features and the true high-frequency features.
[0148] ii. In on example, the cosine similarity loss can be used to constrain the difference between the predicted high-frequency features and the true high-frequency features.
[0149] iii. In on example, the adversarial loss is used to constrain the distribution of predicted high frequency features.
[0150] 6) It is proposed to get the reconstructed point cloud based on the up-sample result of the decoded point cloud.
[0151] a. In one example, the reconstructed point cloud may be obtained directly by up-sampling in one time.
[0152] b. In one example, the reconstructed point cloud may be obtained directly by multiple progressive up-sampling.
[0153] i. In one example, the point cloud may be reconstructed using N up-sampling operations, such as N = 3.
[0154] 1. In one example, N may be pre-defined.
[0155] 2. In one example, N may be signalled to the decoder.
[0156] a. In one example, N may be coded with fixed-length coding, unary coding, truncated unary coding, etc. al.
[0157] b. In one example, N may be coding in a predictive way.
[0158] c. In one example, the sparse convolution-based generative convolution may be used to achieve point cloud up-sampling.
[0159] d. In one example, multi-stage loss functions with different granularities may be used to constrain neural network during the training process of the up-sampling.
[0160] i. In one example, the binary cross-entropy value may be used as the loss function in the first stage.
[0161] ii. In one example, the numbers of points used in the loss function may be different in different stages.
[0162] iii. In one example, the number of points used in the loss function for each staged may be indicated by at least one indication.
[0163] 1. In one example, there are 3 stages and 3 indications, M, N and K.
[0164] a. In one example, M%points of point cloud obtained by voxel sampling from the real point cloud may be used to constrain the reconstructed point cloud in the first stage.
[0165] b. In one example, N%points of point cloud obtained by voxel sampling from the real point cloud may be used to constrain the reconstructed point cloud in the first stage.
[0166] c. In one example, K%points of point cloud obtained by voxel sampling from the real point cloud may be used to constrain the reconstructed point cloud in the last stage.
[0167] d. In one example, M < N < K, such as M = 12.5, N = 50, K = 100.
[0168] iv. In one example, the indication may be pre-defined.
[0169] v. In one example, the indication may be signalled.
[0170] 7) It is proposed to train the downsampling and upsampling method in different training modes.
[0171] a. In one example, downsampling and upsampling may be trained simultaneously.
[0172] b. In one example, downsampling and upsampling may be trained separately.
[0173] c. In one example, downsampling and upsampling may be trained at the same time, and then up-sampling and down-sampling may be trained separately.
[0174] 8) Whether to and / or how to apply a method disclosed above may be signaled from encoder to decoder in a bitstream / frame / tile / slice / octree / etc.
[0175] 9) Whether to and / or how to apply the disclosed methods above may be dependent on coded information, such as dimensions, colour format, colour component, slice / picture type.
[0176] 6. Embodiments
[0177] An example of the coding flow for the collaborative adaptive downsampling and upsampling method of point clouds is depicted in Fig. 4. An example of the flow for the collaborative adaptive downsampling and upsampling method of point clouds is as follows. Firstly, the high- resolution original point cloud is passed through a learnable adaptive downsampling network to obtain a low-resolution point cloud that best fits the up-recovery network, which greatly reduces the amount of data in the original point cloud and thus the bit rate. Secondly, the initial features are assigned to the low-resolution point cloud, after the low-resolution point cloud with basic quality is solved from the decoder. Thirdly, the feature prediction module can generate the high frequency information features lost in the downsampling process, which can better guide the point cloud upsampling recovery. Last, the point clouds are reconstructed using progressive upsampling to reduce the reconstruction difficulty, and the reconstruction results are constrained using multi-stage loss function Thus, the multi-stage loss constraint could make the reconstruction better for overall structure.
[0178] More details of the embodiments of the present disclosure will be described below which are related to collaborative adaptive downsampling and upsampling of point clouds. The embodiments of the present disclosure should be considered as examples to explain the general concepts and should not be interpreted in a narrow way. Furthermore, these embodiments can be applied individually or combined in any manner.
[0179] As used herein, the term “point cloud sequence” may refer to a sequence of one or more point clouds. The term “point cloud frame” or “frame” may refer to a point cloud in a point cloud sequence. The term “point cloud (PC) sample” may refer to a frame, a sub-region within a frame, a picture, a slice, a sub-frame, a sub-picture, a tile, a segment, or any other suitable processing unit.
[0180] Fig. 5 illustrates a flowchart of a method 500 for point cloud processing in accordance with some embodiments of the present disclosure. The method 500 may be implemented during a conversion between a current PC sample of a point cloud sequence and a bitstream of the point cloud sequence. At 502, a first set of points in the current PC sample is obtained. The first set of points representing low frequency information of the current PC sample. For example, the first set of points may be regard as a sparse point cloud, and low frequency information of the current PC sample may be determined based on the sparse point cloud.
[0181] In some embodiments, in the encoding process, the first set of points may be obtained by downsampling the current PC sample, and signaled the bitstream. Accordingly, in the decoding process, the first set of points may be obtained from the bitstream, e.g. by parsing the bitstream. This will be described in detail below.
[0182] At 504, the conversion is performed based on the first set of points. In some embodiments the conversion may include encoding the current PC sample into the bitstream. Alternatively or additionally, the conversion may include decoding the current PC sample from the bitstream.
[0183] By way of example rather than limitation, at 504, a first feature associated with the current PC sample is generated based on the first set of points. The first feature may be associated with high frequency information of the current PC sample, and may also be referred to as “initial feature” . For example, the first feature may be generated based on the first set of points by using an ML-based model. In some example embodiments, the first feature may be signaled in the bitstream. Alternatively, the first feature may be absent from the bitstream.
[0184] Furthermore, a prediction for a high frequency feature may be generated based on the first feature. The high frequency feature may represent high frequency information of the current PC sample. The conversion may be performed based on the prediction for the high frequency feature. For example, the current PC sample may be reconstructed by applying a upsampling process on the first set of points based on the prediction for the high frequency feature. It should be understood that the above illustrations are described merely for purpose of description. The scope of the present disclosure is not limited in this respect.
[0185] In view of the above, a point cloud is coded based on a set of points representing low frequency information of the point cloud. Compared with the conventional solution, the proposed method can advantageously utilize collaborative adaptive downsampling and upsampling of a point cloud to assist in coding point cloud. Thereby, the coding efficiency can be improved.
[0186] In some embodiments, the prediction for the high frequency feature may be generated with a machine learning based (ML-based) model. As used herein, the term “model” is referred to as an association between an input and an output learned from training data, and thus a corresponding output may be generated for a given input after the training. The generation of the model may be based on a machine learning technique. In general, a machine learning model may be built, which receives input information and makes predictions based on the input information. For example, a classification model may predict a class of the input information among a predetermined set of classes. As used herein, “model” may also be referred to as “machine learning model” , “learning model” , “machine learning network” , or “learning network, ” which are used interchangeably herein.
[0187] For example, the ML-based model may comprise a convolution operation. Additionally or alternatively, the ML-based model may comprise a general feature extraction network. In some further example embodiments, the ML-based model may comprise a generative adversarial network, and / or the like.
[0188] In some embodiments, a generator in the generative adversarial network may comprise a convolutional block with feature expansion operations. For example, the convolutional block may be implemented with at least one sparse convolution. By way of example rather than limitation, the at least one sparse convolution may comprise a plurality of consecutive sparse convolutions, such as three or five consecutive sparse convolutions. It should be understood that the specific values recited herein are intended to be exemplary rather than limiting the scope of the present disclosure.
[0189] In some additional or alternative embodiments, a discriminator in the generative adversarial network may comprise a convolutional block. For example, the convolutional block may be implemented with at least one sparse convolution. By way of example rather than limitation, the at least one sparse convolution may comprise a plurality of consecutive sparse convolutions, such as three or four consecutive sparse convolutions. In addition, the discriminator may further comprise a maximum pooling operation, so as to retain the detail information. Additionally or alternatively, the discriminator may further comprise an average pooling operation, so as to retain the overall structure information.
[0190] In some embodiments, at least one parameter of the ML-based model may be updated based on an adversarial loss function during a training process of the ML-based model. It should be noted that any other suitable loss function (such as L2 distance loss function, cosine similarity loss or the like) may be employed to train the ML-based model.
[0191] In some embodiments, a second feature associated with local high frequency information for a current point of the current PC sample may be generated. Moreover, a third feature associated with global high frequency information for the current PC sample may be generated. As used herein, the second feature may also be referred to as “local high frequency feature” , and the third feature may also be referred to as “global high frequency feature” .
[0192] For example, the second feature and / or the third feature may be generated based on a second set of points obtained by downsampling the current PC sample. In some example embodiments, the second set of points may be the same as the first set of points. Alternatively, the second set of points may be different from the first set of points.
[0193] In some embodiments, the second feature may be generated with a neural network. For example, the second feature may be generated with a first graph convolution. As used herein, graph convolution may also be referred to as graph aggregation.
[0194] In some embodiments, a first graph (which may also be referred to as “local graph” ) used for the first graph convolution may be generated based on a set of nearest neighbors of the current point, e.g., based on the K-Nearest Neighbors (KNN) . In some embodiments, the set of nearest neighbors may be determined based on a distance in Euclidean space or a distance in feature space. As used herein, a feature of a point may indicate occupancy information, a specific distance, or a value derived based on geometric information.
[0195] In some embodiments, in the first graph convolution, features within the first graph may be aggregated by determining a sum of features of all points in the first graph. Alternatively, in the first graph convolution, features within the first graph may be aggregated by determining a weighted sum of features of all points in the first graph. In this case, a weight for weighting a feature of a first point in the first graph may be dependent on a geometric distance between the first point and the current point. For example, the larger the distance between the first point and the current point, the higher the weight assigned to the feature of the first point. Alternatively, the smaller the distance between the first point and the current point, the higher the weight assigned to the feature of the first point. In some further example embodiments, a weight for weighting a feature of a second point in the first graph may be dependent on a feature similarity between the second point and the current point. In this case, the higher the similarity between features of the second point and the current point, the higher the weight assigned to the feature of the second point. Alternatively, the lower the similarity between features of the second point and the current point, the higher the weight assigned to the feature of the second point.
[0196] In some embodiments, the first graph convolution may be implemented based on a sparse convolution. For example, a step size of the sparse convolution may be equal to a first number, such as 2, 3, or the like. In one example embodiment, the first number may be predetermined. In a further example embodiment, the first number may be indicated in the bitstream. In some embodiments, the second feature may be determined as a feature obtained by convolutional downsampling.
[0197] In some embodiments, the third feature may be generated with a neural network. For example, the third feature may be generated with a second graph convolution. In some embodiments, a second graph (which may also be referred to as “global graph” ) used for the second graph convolution may be generated based on all points of the current PC sample.
[0198] In one example embodiment, the second graph may be generated with a non-learning-based scheme. By way of example rather than limitation, the non-learning-based scheme may comprise farthest point sampling or the like. In a further example embodiment, the second graph may be generated with a learning-based scheme. For example, in the learning-based scheme, a response value may be generated for each point in the current PC sample by using an ML-based model. Then, one or more points for generating the second graph may be selected from all points of the current PC sample based on the respective response values. By way of example rather than limitation, the plurality of points may comprise points with top N response values, and N may be an integer, such as 8, 12, or the like.
[0199] In some embodiments, in the second graph convolution, features within the second graph may be aggregated by determining a sum of features of all points in the second graph. In some alternative embodiments, in the second graph convolution, features within the second graph may be aggregated by determining a weighted sum of features of all points in the second graph. In this case, a weight for weighting a feature of a third point in the second graph may be dependent on a geometric distance between the third point and the current point. For example, the larger the distance between the third point and the current point, the higher the weight assigned to the feature of the third point. Alternatively, the smaller the distance between the third point and the current point, the higher the weight assigned to the feature of the third point. In some further embodiments, a weight for weighting a feature of a fourth point in the second graph may be dependent on a feature similarity between the fourth point and the current point. In this case, the higher the similarity between features of the fourth point and the current point, the higher the weight assigned to the feature of the fourth point. Alternatively, the lower the similarity between features of the fourth point and the current point, the higher the weight assigned to the feature of the fourth point.
[0200] In some embodiments, the second graph convolution may be implemented based on a sparse convolution. For example, a step size of the sparse convolution may be equal to a second number, such as 2, 4 or the like. In one example embodiment, the second number may be predetermined. Alternatively, the second number may be indicated in the bitstream. In some embodiments, the third feature may be determined as a feature obtained by convolutional downsampling.
[0201] In some embodiments, an ML-based model for generating the prediction for the high frequency feature may be trained based on the second feature and / or the third feature. For example, in aid of the second feature and the third feature, a supervised training process may be applied on the ML-based model. Thereby, the trained ML-based model can better predict the high frequency feature, and thus the coding quality can be improved.
[0202] In view of the above, the solutions in accordance with some embodiments of the present disclosure can advantageously improve coding performance.
[0203] According to further embodiments of the present disclosure, a non-transitory computer-readable recording medium is provided. The non-transitory computer-readable recording medium stores a bitstream of a point cloud sequence which is generated by a method performed by an apparatus for point cloud processing. In the method, a first set of points in a current PC sample of the point cloud sequence is obtained. The first set of points represents low frequency information of the current PC sample. Moreover, the bitstream is generated based on the first set of points.
[0204] According to still further embodiments of the present disclosure, a method for storing bitstream of a point cloud sequence is provided. In the method, a first set of points in a current PC sample of the point cloud sequence is obtained. The first set of points represents low frequency information of the current PC sample. Moreover, the bitstream is generated based on the first set of points, and stored in a non-transitory computer-readable recording medium.
[0205] Implementations of the present disclosure can be described in view of the following clauses, the features of which can be combined in any reasonable manner.
[0206] Clause 1. A method for point cloud processing, comprising: obtaining, for a conversion between a current point cloud (PC) sample of a point cloud sequence and a bitstream of the point cloud sequence, a first set of points in the current PC sample, the first set of points representing low frequency information of the current PC sample; and performing the conversion based on the first set of points.
[0207] Clause 2. The method of clause 1, wherein performing the conversion comprises: generating, based on the first set of points, a first feature associated with the current PC sample; generating, based on the first feature, a prediction for a high frequency feature, the high frequency feature representing high frequency information of the current PC sample; and performing the conversion based on the prediction for the high frequency feature.
[0208] Clause 3. The method of clause 2, wherein performing the conversion based on the prediction for the high frequency feature comprises: reconstructing the current PC sample by applying a upsampling process on the first set of points based on the prediction for the high frequency feature.
[0209] Clause 4. The method of any of clauses 2-3, wherein the prediction for the high frequency feature is generated with a machine learning based (ML-based) model comprising at least one of the following: a convolution operation, a general feature extraction network, or a generative adversarial network.
[0210] Clause 5. The method of clause 4, wherein a generator in the generative adversarial network comprises a convolutional block with feature expansion operations.
[0211] Clause 6. The method of clause 5, wherein the convolutional block is implemented with at least one sparse convolution.
[0212] Clause 7. The method of clause 6, wherein the at least one sparse convolution comprises a plurality of consecutive sparse convolutions.
[0213] Clause 8. The method of any of clauses 4-7, wherein a discriminator in the generative adversarial network comprises at least one of the following: a convolutional block, a maximum pooling operation, or an average pooling operation.
[0214] Clause 9. The method of clause 8, wherein the convolutional block is implemented with at least one sparse convolution.
[0215] Clause 10. The method of clause 9, wherein the at least one sparse convolution comprises a plurality of consecutive sparse convolutions.
[0216] Clause 11. The method of any of clauses 4-10, wherein at least one parameter of the ML-based model is updated based on an adversarial loss function during a training process of the ML-based model.
[0217] Clause 12. The method of any of clauses 2-11, wherein the first feature is generated based on the first set of points by using an ML-based model.
[0218] Clause 13. The method of any of clauses 1-12, further comprising: generating a second feature associated with local high frequency information for a current point of the current PC sample; and generating a third feature associated with global high frequency information for the current PC sample.
[0219] Clause 14. The method of clause 13, wherein at least one of the second feature or the third feature is generated based on a second set of points obtained by downsampling the current PC sample.
[0220] Clause 15. The method of any of clauses 13-14, wherein the second feature is generated with a neural network.
[0221] Clause 16. The method of any of clauses 13-15, wherein the second feature is generated with a first graph convolution.
[0222] Clause 17. The method of clause 16, wherein a first graph used for the first graph convolution is generated based on a set of nearest neighbors of the current point.
[0223] Clause 18. The method of clause 17, wherein the set of nearest neighbors is determined based on a distance in Euclidean space or a distance in feature space.
[0224] Clause 19. The method of any of clauses 17-18, wherein in the first graph convolution, features within the first graph are aggregated by determining a sum of features of all points in the first graph.
[0225] Clause 20. The method of any of clauses 17-18, wherein in the first graph convolution, features within the first graph are aggregated by determining a weighted sum of features of all points in the first graph.
[0226] Clause 21. The method of clause 20, wherein a weight for weighting a feature of a first point in the first graph is dependent on a geometric distance between the first point and the current point.
[0227] Clause 22. The method of clause 20, wherein a weight for weighting a feature of a second point in the first graph is dependent on a feature similarity between the second point and the current point.
[0228] Clause 23. The method of any of clauses 16-22, wherein the first graph convolution is implemented based on a sparse convolution.
[0229] Clause 24. The method of clause 23, wherein a step size of the sparse convolution is equal to a first number.
[0230] Clause 25. The method of clause 24, wherein the first number is predetermined or indicated in the bitstream.
[0231] Clause 26. The method of any of clauses 13-25, wherein the second feature is determined as a feature obtained by convolutional downsampling.
[0232] Clause 27. The method of any of clauses 13-26, wherein the third feature is generated with a neural network.
[0233] Clause 28. The method of any of clauses 13-27, wherein the third feature is generated with a second graph convolution.
[0234] Clause 29. The method of clause 28, wherein a second graph used for the second graph convolution is generated based on all points of the current PC sample.
[0235] Clause 30. The method of clause 29, wherein the second graph is generated with a non-learning-based scheme.
[0236] Clause 31. The method of clause 30, wherein the non-learning-based scheme comprises farthest point sampling.
[0237] Clause 32. The method of clause 29, wherein the second graph is generated with a learning-based scheme.
[0238] Clause 33. The method of clause 32, wherein in the learning-based scheme, a response value is generated for each point in the current PC sample by using an ML-based model, and a plurality of points for generating the second graph are selected from all points of the current PC sample based on the respective response values.
[0239] Clause 34. The method of clause 33, wherein the plurality of points comprises points with top N response values, and N is an integer.
[0240] Clause 35. The method of clause 34, wherein N is equal to 8.
[0241] Clause 36. The method of any of clauses 29-35, wherein in the second graph convolution, features within the second graph are aggregated by determining a sum of features of all points in the second graph.
[0242] Clause 37. The method of any of clauses 29-35, wherein in the second graph convolution, features within the second graph are aggregated by determining a weighted sum of features of all points in the second graph.
[0243] Clause 38. The method of clause 37, wherein a weight for weighting a feature of a third point in the second graph is dependent on a geometric distance between the third point and the current point.
[0244] Clause 39. The method of clause 37, wherein a weight for weighting a feature of a fourth point in the second graph is dependent on a feature similarity between the fourth point and the current point.
[0245] Clause 40. The method of any of clauses 28-39, wherein the second graph convolution is implemented based on a sparse convolution.
[0246] Clause 41. The method of clause 40, wherein a step size of the sparse convolution is equal to a second number.
[0247] Clause 42. The method of clause 41, wherein the second number is predetermined or indicated in the bitstream.
[0248] Clause 43. The method of any of clauses 13-42, wherein the third feature is determined as a feature obtained by convolutional downsampling.
[0249] Clause 44. The method of any of clauses 13-43, wherein an ML-based model for generating the prediction for the high frequency feature is trained based on at least one of the second feature or the third feature.
[0250] Clause 45. The method of any of clauses 1-44, wherein the first set of points is obtained from the bitstream.
[0251] Clause 46. The method of any of clauses 1-44, wherein the first set of points is obtained by downsampling the current PC sample.
[0252] Clause 47. The method of any of clauses 1-46, wherein a PC sample is one of the following: a frame, a picture, a slice, a sub-frame, a sub-picture, a tile, or a segment.
[0253] Clause 48. The method of any of clauses 1-47, wherein the conversion includes encoding the current PC sample into the bitstream.
[0254] Clause 49. The method of any of clauses 1-47, wherein the conversion includes decoding the current PC sample from the bitstream.
[0255] Clause 50. An apparatus for point cloud processing comprising a processor and a non-transitory memory with instructions thereon, wherein the instructions upon execution by the processor, cause the processor to perform a method in accordance with any of clauses 1-49.
[0256] Clause 51. A non-transitory computer-readable storage medium storing instructions that cause a processor to perform a method in accordance with any of clauses 1-49.
[0257] Clause 52. A non-transitory computer-readable recording medium storing a bitstream of a point cloud sequence which is generated by a method performed by an apparatus for point cloud processing, wherein the method comprises: obtaining a first set of points in a current PC sample of the point cloud sequence, the first set of points representing low frequency information of the current PC sample; and generating the bitstream based on the first set of points.
[0258] Clause 53. A method for storing a bitstream of a point cloud sequence, comprising: obtaining a first set of points in a current PC sample of the point cloud sequence, the first set of points representing low frequency information of the current PC sample; generating the bitstream based on the first set of points; and storing the bitstream in a non-transitory computer-readable recording medium.
[0259] Example Device
[0260] Fig. 6 illustrates a block diagram of a computing device 600 in which various embodiments of the present disclosure can be implemented. The computing device 600 may be implemented as or included in the source device 110 (or the GPCC encoder 116 or 200) or the destination device 120 (or the GPCC decoder 126 or 300) .
[0261] It would be appreciated that the computing device 600 shown in Fig. 6 is merely for purpose of illustration, without suggesting any limitation to the functions and scopes of the embodiments of the present disclosure in any manner.
[0262] As shown in Fig. 6, the computing device 600 includes a general-purpose computing device 600. The computing device 600 may at least comprise one or more processors or processing units 610, a memory 620, a storage unit 630, one or more communication units 640, one or more input devices 650, and one or more output devices 660.
[0263] In some embodiments, the computing device 600 may be implemented as any user terminal or server terminal having the computing capability. The server terminal may be a server, a large-scale computing device or the like that is provided by a service provider. The user terminal may for example be any type of mobile terminal, fixed terminal, or portable terminal, including a mobile phone, station, unit, device, multimedia computer, multimedia tablet, Internet node, communicator, desktop computer, laptop computer, notebook computer, netbook computer, tablet computer, personal communication system (PCS) device, personal navigation device, personal digital assistant (PDA) , audio / video player, digital camera / video camera, positioning device, television receiver, radio broadcast receiver, E-book device, gaming device, or any combination thereof, including the accessories and peripherals of these devices, or any combination thereof. It would be contemplated that the computing device 600 can support any type of interface to a user (such as “wearable” circuitry and the like) .
[0264] The processing unit 610 may be a physical or virtual processor and can implement various processes based on programs stored in the memory 620. In a multi-processor system, multiple processing units execute computer executable instructions in parallel so as to improve the parallel processing capability of the computing device 600. The processing unit 610 may also be referred to as a central processing unit (CPU) , a microprocessor, a controller or a microcontroller.
[0265] The computing device 600 typically includes various computer storage medium. Such medium can be any medium accessible by the computing device 600, including, but not limited to, volatile and non-volatile medium, or detachable and non-detachable medium. The memory 620 can be a volatile memory (for example, a register, cache, Random Access Memory (RAM) ) , a non-volatile memory (such as a Read-Only Memory (ROM) , Electrically Erasable Programmable Read-Only Memory (EEPROM) , or a flash memory) , or any combination thereof. The storage unit 630 may be any detachable or non-detachable medium and may include a machine-readable medium such as a memory, flash memory drive, magnetic disk or another other media, which can be used for storing information and / or data and can be accessed in the computing device 600.
[0266] The computing device 600 may further include additional detachable / non-detachable, volatile / non-volatile memory medium. Although not shown in Fig. 6, it is possible to provide a magnetic disk drive for reading from and / or writing into a detachable and non-volatile magnetic disk and an optical disk drive for reading from and / or writing into a detachable non-volatile optical disk. In such cases, each drive may be connected to a bus (not shown) via one or more data medium interfaces.
[0267] The communication unit 640 communicates with a further computing device via the communication medium. In addition, the functions of the components in the computing device 600 can be implemented by a single computing cluster or multiple computing machines that can communicate via communication connections. Therefore, the computing device 600 can operate in a networked environment using a logical connection with one or more other servers, networked personal computers (PCs) or further general network nodes.
[0268] The input device 650 may be one or more of a variety of input devices, such as a mouse, keyboard, tracking ball, voice-input device, and the like. The output device 660 may be one or more of a variety of output devices, such as a display, loudspeaker, printer, and the like. By means of the communication unit 640, the computing device 600 can further communicate with one or more external devices (not shown) such as the storage devices and display device, with one or more devices enabling the user to interact with the computing device 600, or any devices (such as a network card, a modem and the like) enabling the computing device 600 to communicate with one or more other computing devices, if required. Such communication can be performed via input / output (I / O) interfaces (not shown) .
[0269] In some embodiments, instead of being integrated in a single device, some or all components of the computing device 600 may also be arranged in cloud computing architecture. In the cloud computing architecture, the components may be provided remotely and work together to implement the functionalities described in the present disclosure. In some embodiments, cloud computing provides computing, software, data access and storage service, which will not require end users to be aware of the physical locations or configurations of the systems or hardware providing these services. In various embodiments, the cloud computing provides the services via a wide area network (such as Internet) using suitable protocols. For example, a cloud computing provider provides applications over the wide area network, which can be accessed through a web browser or any other computing components. The software or components of the cloud computing architecture and corresponding data may be stored on a server at a remote position. The computing resources in the cloud computing environment may be merged or distributed at locations in a remote data center. Cloud computing infrastructures may provide the services through a shared data center, though they behave as a single access point for the users. Therefore, the cloud computing architectures may be used to provide the components and functionalities described herein from a service provider at a remote location. Alternatively, they may be provided from a conventional server or installed directly or otherwise on a client device.
[0270] The computing device 600 may be used to implement point cloud encoding / decoding in embodiments of the present disclosure. The memory 620 may include one or more point cloud processing modules 625 having one or more program instructions. These modules are accessible and executable by the processing unit 610 to perform the functionalities of the various embodiments described herein.
[0271] In the example embodiments of performing point cloud encoding, the input device 650 may receive point cloud data as an input 670 to be encoded. The point cloud data may be processed, for example, by the point cloud processing module 625, to generate an encoded bitstream. The encoded bitstream may be provided via the output device 660 as an output 680.
[0272] In the example embodiments of performing point cloud decoding, the input device 650 may receive an encoded bitstream as the input 670. The encoded bitstream may be processed, for example, by the point cloud processing module 625, to generate decoded point cloud data. The decoded point cloud data may be provided via the output device 660 as the output 680.
[0273] While this disclosure has been particularly shown and described with references to preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the present application as defined by the appended claims. Such variations are intended to be covered by the scope of this present application. As such, the foregoing description of embodiments of the present application is not intended to be limiting.
Claims
1.A method for point cloud processing, comprising:obtaining, for a conversion between a current point cloud (PC) sample of a point cloud sequence and a bitstream of the point cloud sequence, a first set of points in the current PC sample, the first set of points representing low frequency information of the current PC sample; andperforming the conversion based on the first set of points.2.The method of claim 1, wherein performing the conversion comprises:generating, based on the first set of points, a first feature associated with the current PC sample;generating, based on the first feature, a prediction for a high frequency feature, the high frequency feature representing high frequency information of the current PC sample; andperforming the conversion based on the prediction for the high frequency feature.3.The method of claim 2, wherein performing the conversion based on the prediction for the high frequency feature comprises:reconstructing the current PC sample by applying a upsampling process on the first set of points based on the prediction for the high frequency feature.4.The method of any of claims 2-3, wherein the prediction for the high frequency feature is generated with a machine learning based (ML-based) model comprising at least one of the following:a convolution operation,a general feature extraction network, ora generative adversarial network.5.The method of claim 4, wherein a generator in the generative adversarial network comprises a convolutional block with feature expansion operations.6.The method of claim 5, wherein the convolutional block is implemented with at least one sparse convolution.7.The method of claim 6, wherein the at least one sparse convolution comprises a plurality of consecutive sparse convolutions.8.The method of any of claims 4-7, wherein a discriminator in the generative adversarial network comprises at least one of the following:a convolutional block,a maximum pooling operation, oran average pooling operation.9.The method of claim 8, wherein the convolutional block is implemented with at least one sparse convolution.10.The method of claim 9, wherein the at least one sparse convolution comprises a plurality of consecutive sparse convolutions.11.The method of any of claims 4-10, wherein at least one parameter of the ML-based model is updated based on an adversarial loss function during a training process of the ML-based model.12.The method of any of claims 2-11, wherein the first feature is generated based on the first set of points by using an ML-based model.13.The method of any of claims 1-12, further comprising:generating a second feature associated with local high frequency information for a current point of the current PC sample; andgenerating a third feature associated with global high frequency information for the current PC sample.14.The method of claim 13, wherein at least one of the second feature or the third feature is generated based on a second set of points obtained by downsampling the current PC sample.15.The method of any of claims 13-14, wherein the second feature is generated with a neural network.16.The method of any of claims 13-15, wherein the second feature is generated with a first graph convolution.17.The method of claim 16, wherein a first graph used for the first graph convolution is generated based on a set of nearest neighbors of the current point.18.The method of claim 17, wherein the set of nearest neighbors is determined based on a distance in Euclidean space or a distance in feature space.19.The method of any of claims 17-18, wherein in the first graph convolution, features within the first graph are aggregated by determining a sum of features of all points in the first graph.20.The method of any of claims 17-18, wherein in the first graph convolution, features within the first graph are aggregated by determining a weighted sum of features of all points in the first graph.21.The method of claim 20, wherein a weight for weighting a feature of a first point in the first graph is dependent on a geometric distance between the first point and the current point.22.The method of claim 20, wherein a weight for weighting a feature of a second point in the first graph is dependent on a feature similarity between the second point and the current point.23.The method of any of claims 16-22, wherein the first graph convolution is implemented based on a sparse convolution.24.The method of claim 23, wherein a step size of the sparse convolution is equal to a first number.25.The method of claim 24, wherein the first number is predetermined or indicated in the bitstream.26.The method of any of claims 13-25, wherein the second feature is determined as a feature obtained by convolutional downsampling.27.The method of any of claims 13-26, wherein the third feature is generated with a neural network.28.The method of any of claims 13-27, wherein the third feature is generated with a second graph convolution.29.The method of claim 28, wherein a second graph used for the second graph convolution is generated based on all points of the current PC sample.30.The method of claim 29, wherein the second graph is generated with a non-learning-based scheme.31.The method of claim 30, wherein the non-learning-based scheme comprises farthest point sampling.32.The method of claim 29, wherein the second graph is generated with a learning-based scheme.33.The method of claim 32, wherein in the learning-based scheme, a response value is generated for each point in the current PC sample by using an ML-based model, and a plurality of points for generating the second graph are selected from all points of the current PC sample based on the respective response values.34.The method of claim 33, wherein the plurality of points comprises points with top N response values, and N is an integer.35.The method of claim 34, wherein N is equal to 8.36.The method of any of claims 29-35, wherein in the second graph convolution, features within the second graph are aggregated by determining a sum of features of all points in the second graph.37.The method of any of claims 29-35, wherein in the second graph convolution, features within the second graph are aggregated by determining a weighted sum of features of all points in the second graph.38.The method of claim 37, wherein a weight for weighting a feature of a third point in the second graph is dependent on a geometric distance between the third point and the current point.39.The method of claim 37, wherein a weight for weighting a feature of a fourth point in the second graph is dependent on a feature similarity between the fourth point and the current point.40.The method of any of claims 28-39, wherein the second graph convolution is implemented based on a sparse convolution.41.The method of claim 40, wherein a step size of the sparse convolution is equal to a second number.42.The method of claim 41, wherein the second number is predetermined or indicated in the bitstream.43.The method of any of claims 13-42, wherein the third feature is determined as a feature obtained by convolutional downsampling.44.The method of any of claims 13-43, wherein an ML-based model for generating the prediction for the high frequency feature is trained based on at least one of the second feature or the third feature.45.The method of any of claims 1-44, wherein the first set of points is obtained from the bitstream.46.The method of any of claims 1-44, wherein the first set of points is obtained by downsampling the current PC sample.47.The method of any of claims 1-46, wherein a PC sample is one of the following:a frame,a picture,a slice,a sub-frame,a sub-picture,a tile, ora segment.48.The method of any of claims 1-47, wherein the conversion includes encoding the current PC sample into the bitstream.49.The method of any of claims 1-47, wherein the conversion includes decoding the current PC sample from the bitstream.50.An apparatus for point cloud processing comprising a processor and a non-transitory memory with instructions thereon, wherein the instructions upon execution by the processor, cause the processor to perform a method in accordance with any of claims 1-49.51.A non-transitory computer-readable storage medium storing instructions that cause a processor to perform a method in accordance with any of claims 1-49.52.A non-transitory computer-readable recording medium storing a bitstream of a point cloud sequence which is generated by a method performed by an apparatus for point cloud processing, wherein the method comprises:obtaining a first set of points in a current PC sample of the point cloud sequence, the first set of points representing low frequency information of the current PC sample; andgenerating the bitstream based on the first set of points.53.A method for storing a bitstream of a point cloud sequence, comprising:obtaining a first set of points in a current PC sample of the point cloud sequence, the first set of points representing low frequency information of the current PC sample;generating the bitstream based on the first set of points; andstoring the bitstream in a non-transitory computer-readable recording medium.
Citation Information
Patent Citations
Point cloud semantic segmentation method and device and electronic equipment
CN110781894A
Three-dimensional point cloud reconstruction result complementing method and related assembly
CN113706686A
Three-dimensional point cloud semantic segmentation method and system based on dynamic aggregation
CN114373104A
Point cloud coding and decoding method and device using adaptive dead zone quantization
CN116547970A
Point cloud decoding and upsampling and model training methods and apparatus
WO2022246724A1