Point cloud compression method, architecture, electronic equipment and storage medium
By integrating the internal geometric features and external view information of point clouds in the point cloud compression framework, and optimizing point cloud compression using encoder and entropy model, the problem of unused view information in the existing technology is solved, and a more efficient point cloud is achieved Compress and rebuild quality.
Patent Information
- Application Number
- CN202410966766.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-18
- Publication Date
- 2024-11-15
AI Technical Summary
The existing point cloud compression framework mainly focuses on the extraction and compression of the internal geometric features of the point cloud, ignores the prior information obtained by the external view, and fails to make full use of this information to improve the compression performance of point cloud.
A point cloud compression method is proposed. By obtaining the original point cloud and the corresponding original view, using an encoder for joint encoding processing, the first point cloud potential representation and the first view potential representation are generated, and entropy encoding is performed through the entropy model. Finally, Reconstruct the point cloud through a decoder, integrating the internal geometric features and external view information of the point cloud.
The reconstruction quality and performance of point cloud compression are improved, while maintaining a high compression ratio, the detail recovery capability of point cloud is enhanced, and the overall performance of point cloud compression is improved.
Smart Images

Figure CN118967848A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data processing technology, and in particular to point cloud compression methods, architectures, electronic devices and storage media. Background Art
[0002] In the field of three-dimensional data processing, point cloud data is an important and widely used form of data representation. Its efficient compression technology is of vital importance for reducing storage requirements, accelerating data transmission and optimizing subsequent processing procedures.
[0003] At present, with the rapid development of deep learning technology, point cloud compression frameworks have gradually become a research hotspot. This type of framework directly analyzes the geometric structure of point cloud data and uses the feature extraction capabilities of deep learning models to achieve efficient encoding and decoding of point cloud data. However, existing point cloud compression frameworks mainly focus on the extraction and compression of internal geometric features of point clouds, but often ignore the prior information obtained from external views. This information is often ignored in current compression frameworks and is not fully utilized.
[0004] In summary, how to effectively integrate the internal geometric features of point clouds and external view information to improve the performance of point cloud compression has become a technical problem that needs to be urgently solved in this field. Summary of the invention
[0005] The main purpose of this application is to provide a point cloud compression method, architecture, electronic device and storage medium, aiming to effectively integrate the internal geometric features of the point cloud and the external view information to improve the performance of point cloud compression.
[0006] To achieve the above objectives, the present application proposes a point cloud compression method, which is applied to a point cloud compression architecture. The architecture includes an encoder, an entropy model and a decoder. The method includes:
[0007] Get the original point cloud and the corresponding original view;
[0008] Performing joint encoding processing on the original point cloud and the original view by the encoder to obtain a first point cloud potential representation and a first view potential representation;
[0009] Performing entropy encoding on the first point cloud potential representation and the first view potential representation by using the entropy model to obtain a second point cloud potential representation and a second view potential representation;
[0010] The second point cloud potential representation and the second view potential representation are decoded by the decoder to obtain a reconstructed point cloud.
[0011] In one embodiment, the encoder includes a first combination module, a residual feature downsampling module, a first modal redundancy removal module, a two-dimensional convolution and a first sparse convolution, wherein the first combination module includes a downsampling unit, a residual network unit and a sparse self-attention unit;
[0012] The step of performing joint encoding processing on the original point cloud and the original view by the encoder to obtain a first point cloud potential representation and a first view potential representation comprises:
[0013] Performing downsampling and feature aggregation processing on the original point cloud through the first combination module to obtain a potential representation of the input point cloud;
[0014] Downsampling and residual feature extraction processing are performed on the original view by the residual feature downsampling module to obtain a potential representation of the input view;
[0015] Performing modal redundancy removal processing on the input point cloud potential representation and the input view potential representation through the first modal redundancy removal module to obtain a third point cloud potential representation and image features;
[0016] Processing the third point cloud potential representation through the first combination module and the first sparse convolution to obtain a first point cloud potential representation;
[0017] The image features are processed by the residual feature downsampling module and the two-dimensional convolution to obtain a first view potential representation.
[0018] In one embodiment, the first modal redundancy removal module includes a sparse cross attention unit, a residual block, a first sparse residual block, a second sparse convolution, a rectified linear unit and a first sigmoid function;
[0019] The step of performing modal redundancy removal processing on the input point cloud potential representation and the input view potential representation by the first modal redundancy removal module to obtain a third point cloud potential representation and image features comprises:
[0020] Extracting features of the potential representation of the input point cloud by using the first sparse residual block to obtain point cloud features;
[0021] Extracting features of the potential representation of the input view by using the residual block to obtain image features;
[0022] Performing modal interaction processing on the point cloud features and the image features through the sparse cross attention unit to obtain an output point cloud potential representation;
[0023] Connecting the output point cloud potential representation and the point cloud feature to obtain a fourth point cloud potential representation;
[0024] Mapping the fourth point cloud potential representation to a preset potential mask to obtain first modal information, wherein the potential mask is the fourth point cloud potential representation after the second sparse convolution, the rectified linear unit and the first sigmoid function;
[0025] The first modality information is removed from the input point cloud latent representation to obtain a third point cloud latent representation.
[0026] In one embodiment, the sparse cross attention unit includes a second sparse residual block, a third sparse convolution, a feature mapping unit, a fully connected layer, and a second sigmoid function;
[0027] The step of performing modal interaction processing on the point cloud features and the image features through the sparse cross attention unit to obtain a potential representation of the output point cloud comprises:
[0028] Processing the point cloud features through the second sparse residual block and the third sparse convolution to obtain a fifth point cloud potential representation;
[0029] Performing image mapping processing on the image features through the feature mapping unit, the fully connected layer and the second Sigmoid function to obtain a mask;
[0030] Mapping the fifth point cloud potential representation to the mask to obtain second modality information;
[0031] The point cloud features and the second modality information are added to obtain an output point cloud potential representation.
[0032] In one embodiment, the entropy model includes a quantizer, an octree encoder, an octree decoder, a conditional checkerboard entropy model, an arithmetic encoder, and an arithmetic decoder;
[0033] The step of entropy encoding the first point cloud potential representation and the first view potential representation by using the entropy model to obtain a second point cloud potential representation and a second view potential representation comprises:
[0034] Splitting the first point cloud potential representation into first point cloud coordinates and first point cloud features;
[0035] Performing quantization operations on the first point cloud feature and the first view potential representation respectively by the quantizer to obtain a point cloud feature quantized potential representation and a view feature quantized potential representation;
[0036] Compressing the first point cloud coordinates by the octree encoder and the octree decoder to obtain point cloud compressed coordinates;
[0037] Performing symbol probability estimation on the point cloud feature quantization potential representation and the view feature quantization potential representation through the conditional checkerboard entropy model to obtain a view entropy parameter, a point cloud anchor point entropy parameter, and a point cloud non-anchor point entropy parameter;
[0038] Through the arithmetic encoder and the arithmetic decoder, a second point cloud potential representation and a second view potential representation are obtained based on the point cloud feature quantization potential representation, the view feature quantization potential representation, the point cloud compressed coordinates, the view entropy parameters, the point cloud anchor point entropy parameters and the point cloud non-anchor point entropy parameters.
[0039] In one embodiment, the conditional checkerboard entropy model includes a super arithmetic encoder, a super arithmetic decoder, a super encoder, a super decoder, a potential transformation module, a potential recovery module, a context prediction module and an entropy parameter prediction module;
[0040] The step of performing symbol probability estimation on the point cloud feature quantization potential representation and the view feature quantization potential representation through the conditional checkerboard entropy model to obtain view entropy parameters, point cloud anchor point entropy parameters and point cloud non-anchor point entropy parameters comprises:
[0041] The point cloud feature quantization potential representation and the view feature quantization potential representation are processed by the super encoder, the super arithmetic encoder, the super arithmetic decoder and the super decoder respectively to obtain point cloud space distribution information and view space distribution information;
[0042] Processing the view space distribution information by the entropy parameter prediction module to obtain a view entropy parameter;
[0043] Splitting the point cloud spatial distribution information and the view spatial distribution information into anchor point potential representations and non-anchor point potential representations respectively;
[0044] Performing modal fusion on the anchor point potential representation and the non-anchor point potential representation through the potential transformation module to obtain an anchor point potential representation and a non-anchor point potential representation;
[0045] Processing the anchor point potential representation and the anchor point potential representation of the point cloud spatial distribution information by the entropy parameter prediction module to obtain a point cloud anchor point entropy parameter;
[0046] The non-anchor potential representation and the context anchor potential representation of the point cloud spatial distribution information are processed by the entropy parameter prediction module to obtain point cloud non-anchor entropy parameters, wherein the context anchor potential representation is obtained by processing the point cloud anchor entropy parameters by the potential recovery module and the context prediction module to obtain the context, and connecting the context and the non-anchor potential representation.
[0047] In one embodiment, the decoder includes a second combination module, a residual feature upsampling module, a quality enhancement module and a second modal redundancy removal module, wherein the second combination module includes an upsampling unit, a residual network unit and a sparse self-attention unit;
[0048] The step of decoding the second point cloud potential representation and the second view potential representation by the decoder to obtain a reconstructed point cloud comprises:
[0049] Processing the second point cloud potential representation by the second combination module to obtain an output point cloud;
[0050] Processing the second view potential representation by the residual feature upsampling module to obtain an output view, and performing modal redundancy removal processing on the output point cloud and the output view by the second modal redundancy removal module;
[0051] The quality enhancement module performs quality enhancement processing on the output point cloud and the output view to obtain a reconstructed point cloud.
[0052] In addition, to achieve the above purpose, the present application also proposes a point cloud compression architecture, which includes an encoder, an entropy model and a decoder;
[0053] The encoder is used to perform joint encoding processing on the original point cloud and the corresponding original view to obtain a first point cloud potential representation and a first view potential representation;
[0054] The entropy model is used to perform entropy encoding on the first point cloud potential representation and the first view potential representation to obtain a second point cloud potential representation and a second view potential representation;
[0055] The decoder is used to decode the second point cloud potential representation and the second view potential representation to obtain a reconstructed point cloud.
[0056] In addition, to achieve the above-mentioned purpose, the present application also proposes an electronic device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the point cloud compression method as described above.
[0057] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium. A computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the point cloud compression method as described above are implemented.
[0058] The present application proposes a point cloud compression method, which is applied to a point cloud compression architecture. The architecture includes an encoder, an entropy model and a decoder. The present application first obtains the original point cloud to be processed and its corresponding original view, and then uses the encoder to jointly encode the original point cloud and the original view to generate a first point cloud potential representation and a first view potential representation. Then, the first point cloud potential representation and the first view potential representation are input into the entropy model for entropy encoding to obtain a second point cloud potential representation and a second view potential representation. Finally, the second point cloud potential representation and the second view potential representation are received by the decoder to obtain a reconstructed point cloud.
[0059] In summary, this application effectively integrates the internal geometric features of the original point cloud with the prior information of the original view corresponding to the device by applying a point cloud compression architecture. Through the joint encoding and entropy model optimization of the two, the detail recovery capability of the point cloud is enhanced, and the reconstruction quality of the point cloud compression is improved, thereby improving the performance of point cloud compression while maintaining a high compression ratio. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.
[0061] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.
[0062] Figure 1 A schematic diagram of a flow chart provided for the first embodiment of the point cloud compression method of the present application;
[0063] Figure 2 A schematic diagram of a point cloud compression architecture provided for Example 2 of the point cloud compression method of the present application;
[0064] Figure 3 A schematic diagram of the structure of a first modal redundancy removal module provided in Embodiment 2 of the point cloud compression method of the present application;
[0065] Figure 4 A schematic diagram of the structure of a sparse cross attention unit provided in Example 2 of the point cloud compression method of the present application;
[0066] Figure 5 A schematic diagram of the structure of the conditional checkerboard entropy model provided in Example 2 of the point cloud compression method of the present application;
[0067] Figure 6 A schematic diagram of the structure of a quality enhancement module provided in Example 2 of the point cloud compression method of the present application;
[0068] Figure 7 A schematic diagram of experimental results provided for Example 2 of the point cloud compression method of this application;
[0069] Figure 8 Another schematic diagram of test results provided for Example 2 of the point cloud compression method of the present application;
[0070] Fig. 9 Schematic diagram of the device structure of the hardware operating environment involved in the point cloud compression method in the embodiment of the present application.
[0071] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0072] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.
[0073] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.
[0074] The main solution of the embodiment of the present application is: applied to a point cloud compression architecture, the architecture including an encoder, an entropy model and a decoder; obtaining an original point cloud and a corresponding original view; performing joint encoding processing on the original point cloud and the original view through the encoder to obtain a first point cloud potential representation and a first view potential representation; performing entropy encoding on the first point cloud potential representation and the first view potential representation through the entropy model to obtain a second point cloud potential representation and a second view potential representation; decoding the second point cloud potential representation and the second view potential representation through the decoder to obtain a reconstructed point cloud.
[0075] The current point cloud compression framework directly analyzes the geometric structure of point cloud data and uses the feature extraction capabilities of deep learning models to achieve efficient encoding and decoding of point cloud data. However, the existing point cloud compression framework mainly focuses on the extraction and compression of internal geometric features of point clouds, but often ignores the prior information obtained from external views. This information is often ignored in the current compression framework and is not fully utilized.
[0076] The embodiment of the present application provides a solution, which effectively integrates the internal geometric features of the original point cloud with the prior information of the original view corresponding to the device by applying a point cloud compression architecture, and enhances the detail recovery capability of the point cloud and improves the reconstruction quality of the point cloud compression through joint encoding and entropy model optimization of the two, thereby improving the performance of point cloud compression while maintaining a high compression ratio.
[0077] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a server, a computer, etc., or an electronic device capable of realizing the above functions, etc. The following takes a computer terminal as an example to illustrate this embodiment and the following embodiments.
[0078] Based on this, the present application embodiment provides a point cloud compression method, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the point cloud compression method of the present application.
[0079] In this embodiment, the point cloud compression method is applied to a point cloud compression architecture, the architecture includes an encoder, an entropy model and a decoder, and the point cloud compression method includes steps S10 to S40:
[0080] Step S10, obtaining the original point cloud and the corresponding original view;
[0081] The original point cloud and the corresponding original view are obtained, wherein the original point cloud contains a large number of points and their attributes in the three-dimensional space, and the original view provides two-dimensional image information observed from a specific perspective.
[0082] Step S20, performing joint encoding processing on the original point cloud and the original view by the encoder to obtain a first point cloud potential representation and a first view potential representation;
[0083] The encoder performs joint encoding of the original point cloud and the original view. In this step, the encoder uses a sparse convolution-based framework to downsample and aggregate features of the original point cloud, and uses a local attention module to extract key features from the original view. Through this joint processing, the encoder generates the first point cloud latent representation and the first view latent representation, which are efficient representations of the original point cloud and the original view in a compressed space, retaining key information and reducing data redundancy.
[0084] Step S30, performing entropy coding on the first point cloud potential representation and the first view potential representation by using the entropy model to obtain a second point cloud potential representation and a second view potential representation;
[0085] After obtaining the first point cloud potential representation and the first view potential representation, entropy coding is performed on these potential representations through an entropy model. Entropy coding is a coding method based on information theory, which can further reduce redundant information in the data.
[0086] Specifically, in this embodiment, the entropy model uses the concept of conditional entropy and introduces the side information of the view hyper-prior as context to make the probability estimation of the entropy parameter more accurate. Through this step, a second point cloud potential representation and a second view potential representation are generated, which achieve a higher compression ratio while retaining key information.
[0087] Step S40: decoding the second point cloud potential representation and the second view potential representation by the decoder to obtain a reconstructed point cloud.
[0088] The second point cloud potential representation and the second view potential representation are decoded by a decoder to reconstruct the compressed point cloud data, that is, to reconstruct the original point cloud.
[0089] During the decoding process, the decoder not only utilizes the potential representation of the point cloud itself, but also fully integrates the corresponding view information. Through the decoder's upsampling, feature recovery and quality enhancement processes, the final output is a reconstructed point cloud. The reconstructed point cloud not only utilizes the geometric structure information of the point cloud itself, but also integrates the view information as prior knowledge, thereby significantly improving the compression performance and reconstruction quality of point cloud compression.
[0090] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can be referred to the above introduction, and will not be repeated in the following. Figure 2 As shown, in the encoder Encoder, a first combination module DIS, a residual feature downsampling module RFD, a first modal redundancy removal module MRRM, a two-dimensional convolution Conv and a first sparse convolution Sconv are included, and the first combination module includes a downsampling unit, a residual network unit and a sparse self-attention unit; step S20 of the point cloud compression method of the present application may include steps S201 to S205:
[0091] Step S201, performing downsampling and feature aggregation processing on the original point cloud through the first combination module to obtain a potential representation of the input point cloud;
[0092] The original point cloud x is firstly rectified by the downsampling unit of the first combination module p Downsampling is performed to reduce data redundancy and extract key features. Subsequently, feature aggregation is performed through residual networks and sparse self-attention units to further enhance the original point cloud x p The feature representation capability of the input point cloud is used to obtain the potential representation of the input point cloud, which not only retains the key information of the original point cloud but also reduces the amount of data.
[0093] Step S202, downsampling and residual feature extraction processing are performed on the original view by the residual feature downsampling module to obtain a potential representation of the input view;
[0094] The original view x is downsampled by the residual feature downsampling module v Downsample and use the residual learning mechanism to extract the original view x v The deep features of are used to obtain a latent representation of the input view, which is an efficient and information-rich compressed form of the original view.
[0095] Step S203, performing modal redundancy removal processing on the input point cloud potential representation and the input view potential representation by the first modal redundancy removal module to obtain a third point cloud potential representation and image features;
[0096] The first modal redundancy removal module uses self-attention and cross-attention mechanisms to analyze and eliminate the modal redundancy between the input point cloud potential representation and the input view potential representation. By mining the intrinsic connection between the point cloud and the view, the data is further compressed and the compression efficiency is improved to obtain the third point cloud potential representation and image features. The third point cloud potential representation removes redundant information while retaining the key point cloud features closely related to the view.
[0097] Step S204, processing the third point cloud potential representation through the first combination module and the first sparse convolution to obtain a first point cloud potential representation;
[0098] The third point cloud potential representation is further processed and compressed by the first combination module and the first sparse convolution module. The first combination module enhances the feature representation of the point cloud by downsampling and feature aggregation. The first sparse convolution module optimizes the sparsity of the point cloud by using sparse convolution technology to further improve the compression efficiency and the accuracy of feature extraction, and obtains the first point cloud potential representation y p ,This representation is an efficient, compact and information-rich compressed form of the original point cloud.
[0099] Step S205: Process the image features through the residual feature downsampling module and the two-dimensional convolution to obtain a first view potential representation.
[0100] The image features are processed by the residual feature downsampling module and the two-dimensional convolution to obtain the first view potential representation y v , which is an efficient, compact and information-rich compressed form of the original view.
[0101] In a feasible embodiment, the structural diagram of the first modal redundancy removal module is as follows: Figure 3As shown, the first modal redundancy removal module MRRM includes a sparse cross attention unit SCA, a residual block RB, a first sparse residual block SRB, a second sparse convolution Sconv, a rectified linear unit ReLU, and a first Sigmoid function Sigmoid; step S203 may include steps S2031 to S2036:
[0102] Step S2031, extracting features of the potential representation of the input point cloud by using the first sparse residual block to obtain point cloud features;
[0103] The first sparse residual block is used to represent the input point cloud latently p Perform deep feature extraction to capture the geometric structure and spatial distribution information in point cloud data and generate point cloud features
[0104] Step S2032, extracting features of the latent representation of the input view by using the residual block to obtain image features;
[0105] The input view is potentially represented by the residual block f v Perform feature extraction to obtain image features In order to retain the key feature information of the view, while alleviating the gradient vanishing problem in deep network training, key features such as texture, color and edges in the view can be effectively extracted.
[0106] Step S2033, performing modal interaction processing on the point cloud features and the image features through the sparse cross attention unit to obtain an output point cloud potential representation;
[0107] Point cloud features extracted by sparse cross-attention unit pairs With image features SCA can intelligently identify and fuse the complementary information in the two modal data to generate an output point cloud potential representation f′ that contains both point cloud geometric information and view semantic information. p .
[0108] Step S2034, connecting the output point cloud potential representation and the point cloud feature to obtain a fourth point cloud potential representation;
[0109] The output point cloud of SCA is potentially represented as f′ p The point cloud features obtained in step S2031 Connect to further fuse and enhance feature representation to form a more comprehensive and rich fourth point cloud potential representation
[0110] Step S2035, mapping the fourth point cloud potential representation to a preset potential mask to obtain first modal information, wherein the potential mask is the fourth point cloud potential representation after the second sparse convolution, the rectified linear unit and the first sigmoid function;
[0111] The fourth point cloud potentially represents Mapped to the preset latent mask m p , to identify and extract the fourth point cloud potential representation The specific information that may cause modal redundancy in the first modality is the information of the first modality, where the potential mask m p is the potential representation of the fourth point cloud after the second sparse convolution, rectified linear unit and the first Sigmoid function.
[0112] Step S2036: remove the first modal information from the input point cloud potential representation to obtain a third point cloud potential representation.
[0113] From the input point cloud latent representation f p Remove the first modal information from the optimization to obtain the third point cloud potential representation To remove unnecessary modal redundancy and improve the accuracy and efficiency of subsequent processing tasks.
[0114] The above processing of MRRM can be summarized as follows:
[0115]
[0116] Among them, F P (·) consists of two consecutive sparse convolutional layers. After an in-depth study of the various components of MRRM and their interactions, it can be seen that MRRM enhances the expressiveness of point cloud features by effectively fusion of features and elimination of redundant modal information, which is particularly important in complex point cloud geometry compression tasks.
[0117] In a feasible embodiment, the structural diagram of the sparse cross attention unit SCA is as follows: Figure 4 As shown, the sparse cross attention unit SCA includes a second sparse residual block SRB, a third sparse convolution SConv, a feature mapping unit Reshape, a fully connected layer Linear and a second Sigmoid function Sigmoid; step S2033 may include steps a10 to a40:
[0118] Step a10, processing the point cloud features through the second sparse residual block and the third sparse convolution to obtain a fifth point cloud potential representation;
[0119] The point cloud features initially extracted by the second sparse residual block Deep processing is performed to further capture its internal complex structure and relationship. Subsequently, a higher-level point cloud feature representation, namely the fifth point cloud potential representation, is extracted through the third sparse convolution. Thereby enhancing the expressive power of point cloud features.
[0120] Step a20, performing image mapping processing on the image features through the feature mapping unit, the fully connected layer and the second Sigmoid function to obtain a mask;
[0121] Image features A series of mapping operations are performed to adapt it to the dimensions and characteristics of the point cloud features. Specifically, the image features are first realized through the feature mapping unit. The mapping from the fully connected layer to a two-dimensional tensor matching the feature dimension of the point cloud ensures that the two have the same dimensional basis when interacting. Subsequently, a fully connected layer is used to further transform the mapped image features, and the second Sigmoid function is used to normalize the output of the fully connected layer to the (0,1) interval to generate a mask. The mask Represents the importance or weight distribution of image features for point cloud features.
[0122] Step a30, mapping the fifth point cloud potential representation to the mask to obtain second modal information; mapping the fifth point cloud potential representation obtained in step a10 to the mask to obtain second modal information; Mapped to the mask generated in step a20 In essence, this mapping process is to perform weighted processing on the point cloud features according to the importance or weight of the image features, so as to extract the modal information that is most relevant or complementary to the image features, that is, the second modal information.
[0123] Step a40, adding the point cloud features and the second modal information to obtain an output point cloud potential representation.
[0124] Point cloud features Added to the second modal information extracted in step a30, the output point cloud potential representation f′ is obtained p ,Specifically, by integrating the contribution of image features into the ,point cloud features, the enhancement and enrichment of the point cloud features are ,achieved. The output point cloud latent representation not only ,contains the geometric information of the original point cloud, but also ,incorporates the semantic and contextual information in the original view, ,providing a more comprehensive and accurate input for subsequent ,processing tasks.
[0125] The above processing of SCA can be summarized as follows:
[0126]
[0127] Where σ(·) represents the Sigmoid function, F L (·) Contains consecutive linear layers.
[0128] In a possible embodiment, Figure 2 As shown, the entropy model includes a quantizer Q, an octree encoder OE, an octree decoder OD, a conditional checkerboard entropy model CCEM, an arithmetic encoder AE and an arithmetic decoder AD; step S30 may include steps S301 to S305:
[0129] Step S301, splitting the first point cloud potential representation into first point cloud coordinates and first point cloud features;
[0130] The first point cloud potential representation y p Decomposed into two parts: the first point cloud coordinate y pc (representing the spatial position of the point cloud) and the first point cloud feature y pf (Contains other properties besides position).
[0131] Step S302, performing quantization operations on the first point cloud feature and the first view latent representation respectively by the quantizer to obtain a point cloud feature quantized latent representation and a view feature quantized latent representation;
[0132] The first point cloud feature y is quantized by pf and the first view latent representation y v Perform quantization, where quantization is a process of reducing data accuracy, aiming to sacrifice certain details in exchange for higher compression rate. This step generates a quantized potential representation of point cloud features. and view features to quantize the latent representation Prepare for subsequent entropy coding.
[0133] Step S303, compressing the first point cloud coordinates by the octree encoder and the octree decoder to obtain point cloud compressed coordinates;
[0134] Through the combination of octree encoder and octree decoder, the first point cloud coordinate y pc The compressed coordinates of the point cloud are generated by compression. Among them, the octree is a tree data structure used for space partitioning, which is particularly suitable for representing sparse point cloud data.
[0135] Step S304, performing symbol probability estimation on the point cloud feature quantization potential representation and the view feature quantization potential representation through the conditional checkerboard entropy model to obtain view entropy parameters, point cloud anchor point entropy parameters and point cloud non-anchor point entropy parameters;
[0136] Quantizing latent representations of point cloud features via conditional checkerboard entropy model and view features to quantize the latent representation The symbol probability is estimated to obtain the view entropy parameters, point cloud anchor entropy parameters and point cloud non-anchor entropy parameters, which will be used in the subsequent arithmetic decoding process. CCEM is an entropy coding technology that can predict the probability distribution of symbols based on the context information of the data (such as the characteristics of neighboring points), thereby achieving more efficient data compression.
[0137] Step S305, through the arithmetic encoder and the arithmetic decoder, based on the point cloud feature quantization potential representation, the view feature quantization potential representation, the point cloud compression coordinates, the view entropy parameters, the point cloud anchor point entropy parameters and the point cloud non-anchor point entropy parameters, obtain a second point cloud potential representation and a second view potential representation.
[0138] Quantize latent representation based on point cloud features through arithmetic decoder View features quantize latent representation Point cloud compression coordinates, view entropy parameters, point cloud anchor entropy parameters and point cloud non-anchor entropy parameters are used to reconstruct the second point cloud potential representation and the second view latent representation These representations not only restore the structure of the original point cloud and original view, but also preserve their key information as much as possible while achieving effective data compression.
[0139] In a feasible embodiment, the schematic diagram of the structure of the conditional checkerboard entropy model CCEM is as follows: Figure 5 As shown, the conditional checkerboard entropy model includes a super arithmetic encoder, a super arithmetic decoder, a super encoder, a super decoder, a potential transformation module, a potential recovery module, a context prediction module and an entropy parameter prediction module; step S304 may include steps S3041 to S3046:
[0140] Step S3041, processing the point cloud feature quantization potential representation and the view feature quantization potential representation respectively through the super encoder, the super arithmetic encoder, the super arithmetic decoder and the super decoder to obtain point cloud space distribution information and view space distribution information;
[0141] The point cloud features are quantized and potentially represented by super encoder, super arithmetic encoder, super arithmetic decoder and super decoder respectively. and the second view latent representation Processing to extract point cloud spatial distribution information and view space distribution information
[0142] Step S3042, processing the view space distribution information by the entropy parameter prediction module to obtain a view entropy parameter;
[0143] View space distribution information Input into the entropy parameter prediction module (EPMv), which predicts and generates the view entropy parameter μ based on the spatial distribution characteristics of the view v and These parameters reflect the probability distribution of symbol occurrences in view features.
[0144] Step S3043, dividing the point cloud spatial distribution information and the view spatial distribution information into anchor point potential representation and non-anchor point potential representation respectively;
[0145] In order to process the spatial distribution information of point cloud more finely and view space distribution information The spatial distribution information of the point cloud is further divided into anchor potential representations and non-anchor latent representations Among them, anchor points usually represent key or representative points in the point cloud, while non-anchor points are the remaining points. Similarly, the view space distribution information Also split into anchor potential representation and non-anchor latent representations This split helps to process these two types of points separately in subsequent steps for more efficient compression.
[0146] Step S3044, performing modality fusion on the anchor point potential representation and the non-anchor point potential representation through the potential transformation module to obtain an anchor point potential representation and a non-anchor point potential representation;
[0147] Through the latent transformation module, the anchor point is potentially represented and non-anchor latent representations Perform modal fusion to combine the information of both to generate a richer anchor potential representation and non-anchor latent representations These representations not only contain their respective original information, but also integrate complementary cross-modal information.
[0148] Step S3045, processing the anchor point potential representation and the anchor point potential representation of the point cloud spatial distribution information by the entropy parameter prediction module to obtain a point cloud anchor point entropy parameter;
[0149] Through the entropy parameter prediction module (EPMa), the anchor point potential representation and anchor potential representation of point cloud spatial distribution information Processing to generate point cloud anchor entropy parameters These parameters reflect the probability distribution of symbols in the anchor points.
[0150] Step S3046, the non-anchor point potential representation and the context anchor point potential representation of the point cloud spatial distribution information are processed by the entropy parameter prediction module to obtain point cloud non-anchor point entropy parameters, wherein the context anchor point potential representation is obtained by processing the point cloud anchor point entropy parameters by the potential recovery module and the context prediction module, and the context and the non-anchor point potential representation are connected.
[0151] The point cloud anchor entropy parameters are adjusted by the potential recovery module and the context prediction module. Processing is performed to predict the context ψ of non-anchor points based on the anchor point information p , then, the predicted context ψ p and non-anchor latent representations are connected to form the contextual anchor potential representation.
[0152] The entropy parameter prediction module (EPMn) is used to represent the non-anchor potential representation of the contextual anchor point and the spatial distribution information of the point cloud. Processing to generate point cloud non-anchor point entropy parameters This parameter reflects the probability distribution of symbols in non-anchor points under a given context, further improving the compression efficiency of non-anchor point data.
[0153] For example, Figure 5 As shown, after the encoder stage, the point cloud features are quantized into latent representations and the second view latent representation They are sent to the super encoder and super decoder to obtain spatial distribution information, and the view spatial distribution information is obtained. and point cloud spatial distribution information Next they are split into anchor points and non-anchor points In order to better integrate the edge information of point cloud (i.e. and ) to exchange information, we use and Yes. All side information is fed into LTM to achieve modal fusion. It should be noted that CCEM aims to make the probability density estimation of point cloud symbols more accurate. For view images, due to their low transmission overhead, after obtaining their side information, we directly feed the side information into the entropy parameter prediction module to obtain the entropy parameters and of the Gaussian distribution. Second view potential representation The characteristics are as follows:
[0154]
[0155] Among them, the formula models the probability of the potential representation of the second view, are some feature maps (latent representations), is modeled by a standard Gaussian distribution, It is the EPMv in the figure, which is responsible for estimating the mean and variance of the Gaussian distribution through the input.
[0156] In order to interact with point cloud edge information more effectively, Respectively After careful alignment, we obtain and The latent representations of all anchors and non-anchors are input into the latent transformation module LTM to achieve modality fusion. The fused anchor latent representations are then Forwarded to EPMa to facilitate the estimation of the probability of the anchor point cloud latent representation. Quantifying the anchor point cloud latent representation The derivation is as follows:
[0157]
[0158] The latent recovery module LRM is derived from μ a and Recovery The context prediction module (CPM) uses sparse masked convolution to exploit Neighborhood information predicts the local context ψ p Therefore, the quantized non-anchor point cloud has the potential It is expressed as follows:
[0159]
[0160] In summary, this embodiment realizes slice-level dependency, thereby greatly improving the encoding and decoding speed. In addition, we introduce cross-modal side information into the entropy model for the first time, which will further improve the symbol probability prediction capability of the entropy model.
[0161] In a possible embodiment, Figure 2 As shown, the decoder includes a second combination module UIS, a residual feature upsampling module RFU, a quality enhancement module VG-QEM and a second modal redundancy removal module MRRM, the second combination module includes an upsampling unit, a residual network unit and a sparse self-attention unit; step S40 may include steps S401 to S403:
[0162] Step S401, processing the second point cloud potential representation by the second combination module to obtain an output point cloud;
[0163] The second point cloud is potentially represented by the second combination module The UIS integrates upsampling, residual network and sparse self-attention unit. These components work together to restore the spatial structure and details of the point cloud. The upsampling process aims to increase the density of the point cloud, the residual network is used to capture and compensate for the information that may be lost during the compression process, and the sparse self-attention unit enhances the model's understanding of the internal relationship of the point cloud. After processing by UIS, the output point cloud is initially restored.
[0164] Step S402, processing the second view potential representation by the residual feature upsampling module to obtain an output view, and performing modal redundancy removal processing on the output point cloud and the output view by the second modal redundancy removal module;
[0165] The potential representation of the second view is obtained through the residual feature upsampling module Processing is performed to restore the corresponding output view, wherein RFU utilizes the upsampling technology of the residual features to effectively amplify the view features and restore the details and clarity of the view.
[0166] At the same time, in order to further improve the quality of the reconstruction results, the second modal redundancy removal module MRRM is introduced to perform modal redundancy removal processing on the output point cloud and output view. MRRM can identify and remove the redundant information that may exist between the point cloud and the view, ensuring the accuracy and efficiency of the reconstruction results.
[0167] Step S403: Performing quality enhancement processing on the output point cloud and the output view by the quality enhancement module to obtain a reconstructed point cloud.
[0168] The quality enhancement module is used to perform quality enhancement on the output point cloud and output view after modal redundancy removal to improve the visual quality and overall expressiveness of the reconstructed point cloud, and finally obtain a high-quality reconstructed point cloud.
[0169] In addition, in a feasible embodiment, the schematic diagram of the architecture of the quality enhancement module VG-QEM is as follows: Figure 6 As shown, at the end of the decoder, we will output a point cloud and the output view Feed into VG-QEM to recover the local texture and geometric structure of the point cloud to achieve better reconstruction quality. VG-QEM is coupled with the decoder, which means we can use an end-to-end training strategy to jointly optimize the encoder-decoder network and VG-QEM. The data processing flow of VG-QEM is as follows: Output point cloud and the output view First, the features are extracted separately, and then the image is sent to the continuous dimensional transformation module (including residual block RB and reshape operator Reshape) to aggregate local details and obtain In each level of operation, the point cloud features after the refinement module (including sparse convolution SConv and ReLU) are Cascading in the channel dimension. This operation is conducive to the coordinate reconstruction of the point cloud on the edge details. Finally, through the residual connection, we get the point cloud with enhanced quality.
[0170] During training, our loss function is:
[0171] L=R p +R v +αD p +βD v ,
[0172] Among them, R p and D p denote the bit rate loss and binary cross entropy (BCE) loss of the point cloud, respectively. Similarly, R v and D v denote the bitrate loss and mean square error (MSE) loss of the view image respectively. α, β are the hyperparameters of the optimization paradigm. p , R v , D p , D v It can be expressed as:
[0173]
[0174] Among them, N p and N v Respectively and The symbol length, M represents the number of scales in the decoder, represents the state of a voxel, indicating whether it is occupied or vacant, and Reflects the probability that the voxel is occupied.
[0175] For example, to help understand the implementation process of the point cloud compression method obtained by combining this embodiment with the above-mentioned embodiment 1, please refer to Figure 2 In general, the point cloud compression architecture used in this embodiment is a dual-branch structure design, which integrates a sparse convolution-based framework to compress the original point cloud (i.e., the Input Point Cloud in the figure), and uses the local attention module to extract features from the view corresponding to the point cloud (i.e., the Input View Image in the figure). Figure 2In the paper, the architecture includes an encoder, an entropy model and a decoder. The first combination module DIS is a combination module including downsampling, a residual network and a sparse self-attention module. The second combination module UIS is similar to DIS in structure, but replaces downsampling with an upsampling process. The MRRM in the encoder is the first modality redundancy removal module, and the MRRM in the decoder is the second modality redundancy removal module. Conv and SConv in the encoder represent the two-dimensional convolution of image data and the first sparse convolution of point cloud data, respectively. CCEM represents the conditional checkerboard entropy model, RFD and RFU are residual feature downsampling and residual feature upsampling, VG-QEM represents the visually guided quality enhancement module, Q represents the quantizer, AE and AD are the arithmetic encoder and arithmetic decoder, OE and OD are the octree encoder and octree decoder.
[0176] In the point cloud compression process, the original point cloud and the corresponding view are fed into the encoder, where downsampling and feature aggregation are performed through DIS, and downsampling and residual feature extraction are performed through RFD. Then, the latent representations of the two modalities are input into MRRM for modal interaction, which effectively removes the redundancy in the point cloud modality. After the encoding process, we obtain the first point cloud latent representation y p and the first view latent representation y v Then, y p Divided into the first point cloud coordinate y pc and the first point cloud feature y pf , for further compression. pf and v are all sent to the quantizer, and y pc Octree-based compression is performed through OE and OD. Then the point cloud features are and view features to quantize the latent representation Input CCEM for accurate symbol probability estimation. At the decoder end, and pc Connect It is used for point cloud feature reconstruction. Then, UIS and RFU are used to and Upsampling and feature recovery are performed. Finally, VG-QEM is used to further enhance the geometric structure and finer details of the point cloud to obtain the reconstructed point cloud. We use the symbols E(·), Split(·), Q(·), Cat(·), and D(·) to designate the encoder, split operation, quantizer, concatenation operation, and decoding, respectively. The main reasoning process is abbreviated as follows:
[0177] y p ,yv =E(x p ,x v ),
[0178] y pc ,y pf =Split(y p ),
[0179]
[0180] For example, in order to further reveal the significant advantages of using the above-mentioned point cloud compression structure for point cloud compression in this embodiment compared with the traditional point cloud compression method, the following will be elaborated and illustrated in detail through specific experimental data.
[0181] In the experiment, the point cloud compression method proposed in this embodiment is compared with several point cloud compression methods with good performance (G-PCC, GeoCNNv2, PCGCv1, PCGCv2, SparsePCGC), and D1 PSNR and D2 PSNR are used as evaluation indicators. D1-PSNR and D2-PSNR reflect the spatial position fidelity of the point cloud and the geometric structure fidelity of the point cloud, respectively. 8iVFB, MVUN, and Owlii are used as comparison data sets.
[0182] like Figure 7 The figure shows the rate-distortion curve used for performance comparison. From left to right, they are Soilder's D1 PSNR, Dancer's D1 PSNR, Soilder's D2 PSNR, and Dancer's D2 PSNR. Figure 7 As shown, the point cloud compression method (red line) of this embodiment is significantly better than other algorithms.
[0183] Table 1 quantitatively demonstrates the performance advantages of the point cloud compression method in this embodiment. Overall, the average RD performance improvement of our proposed method over the existing best model is more than 10% and 8% in D1-PSNR and D2-PSNR indicators, respectively. Our method performs best on Owlii, which can further reduce our image bit overhead due to its large number of points.
[0184] Table 1
[0185]
[0186] In order to clearly demonstrate the subjective visual effect of the compressed point cloud of the present invention, we visualize the point clouds compressed by the method proposed in the present invention and several competitive models, such as Figure 8As shown. For the sake of fairness, we tried to keep all models compressed at the same bit rate, and their actual bit rates are 0.27, 0.26, 0.22, 0.21, and 0.19 from left to right in the order shown in the figure. The local details of our point cloud (such as the texture on the clothes) are enlarged to more intuitively show our reconstruction results. It can be clearly seen that G-PCC has obvious blocking effects during reconstruction, while methods such as SparsePCGC have white spots of varying degrees during reconstruction, which indicates that there is geometric distortion in the area. Under the premise of using a lower bit rate, the reconstructed point cloud of our method is closest to the texture features of the original point cloud, which also reflects the superior subjective visual quality of the reconstructed point cloud of our method.
[0187] In this way, the experimental data intuitively illustrates that the point cloud compression method provided in this embodiment can reconstruct a reconstructed point cloud with better visual quality.
[0188] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the point cloud compression method of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.
[0189] The embodiment of the present application also provides a point cloud compression architecture, which includes: an encoder, an entropy model and a decoder;
[0190] The encoder is used to perform joint encoding processing on the original point cloud and the corresponding original view to obtain a first point cloud potential representation and a first view potential representation;
[0191] The entropy model is used to perform entropy encoding on the first point cloud potential representation and the first view potential representation to obtain a second point cloud potential representation and a second view potential representation;
[0192] The decoder is used to decode the second point cloud potential representation and the second view potential representation to obtain a reconstructed point cloud.
[0193] The point cloud compression architecture provided in the embodiment of the present application adopts the point cloud compression method in the above embodiment, which can effectively fuse the internal geometric features of the point cloud and the external view information to improve the performance of point cloud compression. Compared with the prior art, the beneficial effects of the point cloud compression architecture provided in the embodiment of the present application are the same as the beneficial effects of the point cloud compression method provided in the above embodiment, and other technical features in the point cloud compression architecture are the same as the features disclosed in the above embodiment method, which will not be repeated here.
[0194] An embodiment of the present application provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the point cloud compression method in the above-mentioned embodiment 1.
[0195] Reference below Fig. 9 , which shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present application. The electronic devices in the embodiments of the present application may include but are not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Fig. 9 The verification value calculation device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.
[0196] like Fig. 9 As shown, the electronic device may include a processing device 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the electronic device are also stored. The processing device 1001, ROM1002, and RAM1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the electronic device to communicate with other devices wirelessly or by wire to exchange data. Although the figure shows an electronic device with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be implemented or have alternatively.
[0197] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.
[0198] The electronic device provided in the embodiment of the present application adopts the point cloud compression method in the above embodiment, which can effectively fuse the internal geometric features of the point cloud and the external view information to improve the performance of point cloud compression. Compared with the prior art, the beneficial effects of the electronic device provided in the embodiment of the present application are the same as the beneficial effects of the point cloud compression method provided in the above embodiment, and the other technical features in the electronic device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.
[0199] It should be understood that the various parts disclosed in the embodiments of the present application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.
[0200] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
[0201] An embodiment of the present application provides a computer-readable storage medium having computer-readable program instructions (ie, a computer program) stored thereon, and the computer-readable program instructions are used to execute the point cloud compression method in the above embodiment.
[0202] The computer-readable storage medium provided in the embodiment of the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM: Random Access Memory), a read-only memory (ROM: Read Only Memory), an erasable programmable read-only memory (EPROM: Erasable Programmable Read Only Memory or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM: CD-Read Only Memory), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency: Radio Frequency), etc., or any suitable combination of the above.
[0203] The computer-readable storage medium may be included in the electronic device, or may exist independently without being installed in the electronic device.
[0204] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by an electronic device, the electronic device: obtains an original point cloud and a corresponding original view; performs joint encoding processing on the original point cloud and the original view through the encoder to obtain a first point cloud potential representation and a first view potential representation; performs entropy encoding on the first point cloud potential representation and the first view potential representation through the entropy model to obtain a second point cloud potential representation and a second view potential representation; decodes the second point cloud potential representation and the second view potential representation through the decoder to obtain a reconstructed point cloud.
[0205] The computer program code for performing the operation of the embodiment of the present application can be written in one or more programming languages or a combination thereof, and the above-mentioned programming languages include object-oriented programming languages, such as Java, Smalltalk, C++, and also include conventional procedural programming languages, such as "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on the remote computer, or completely on the remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (e.g., using an Internet service provider to connect through the Internet).
[0206] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.
[0207] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.
[0208] The readable storage medium provided in the embodiment of the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned point cloud compression method, and can effectively fuse the internal geometric features of the point cloud and the external view information to improve the performance of point cloud compression. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in the embodiment of the present application are the same as the beneficial effects of the point cloud compression method provided in the above-mentioned embodiment, and will not be repeated here.
[0209] The above are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect application in other related technical fields are included in the patent protection scope of the present application.
Claims
1. A point cloud compression method, It is characterized in that Applied to a point cloud compression architecture, the architecture includes an encoder, an entropy model and a decoder, the method includes: Get the original point cloud and the corresponding original view; Performing joint encoding processing on the original point cloud and the original view by the encoder to obtain a first point cloud potential representation and a first view potential representation; Performing entropy encoding on the first point cloud potential representation and the first view potential representation by using the entropy model to obtain a second point cloud potential representation and a second view potential representation; The second point cloud potential representation and the second view potential representation are decoded by the decoder to obtain a reconstructed point cloud.
2. The method according to claim 1, It is characterized in that The encoder includes a first combination module, a residual feature downsampling module, a first modal redundancy removal module, a two-dimensional convolution and a first sparse convolution, wherein the first combination module includes a downsampling unit, a residual network unit and a sparse self-attention unit; The step of performing joint encoding processing on the original point cloud and the original view by the encoder to obtain a first point cloud potential representation and a first view potential representation comprises: Performing downsampling and feature aggregation processing on the original point cloud through the first combination module to obtain a potential representation of the input point cloud; Downsampling and residual feature extraction processing are performed on the original view by the residual feature downsampling module to obtain a potential representation of the input view; Performing modal redundancy removal processing on the input point cloud potential representation and the input view potential representation through the first modal redundancy removal module to obtain a third point cloud potential representation and image features; Processing the third point cloud potential representation through the first combination module and the first sparse convolution to obtain a first point cloud potential representation; The image features are processed by the residual feature downsampling module and the two-dimensional convolution to obtain a first view potential representation.
3. The method according to claim 2, It is characterized in that The first modal redundancy removal module includes a sparse cross attention unit, a residual block, a first sparse residual block, a second sparse convolution, a rectified linear unit and a first sigmoid function; The step of performing modal redundancy removal processing on the input point cloud potential representation and the input view potential representation by the first modal redundancy removal module to obtain a third point cloud potential representation and image features comprises: Extracting features of the potential representation of the input point cloud by using the first sparse residual block to obtain point cloud features; Extracting features of the potential representation of the input view by using the residual block to obtain image features; Performing modal interaction processing on the point cloud features and the image features through the sparse cross attention unit to obtain an output point cloud potential representation; Connecting the output point cloud potential representation and the point cloud feature to obtain a fourth point cloud potential representation; Mapping the fourth point cloud potential representation to a preset potential mask to obtain first modal information, wherein the potential mask is the fourth point cloud potential representation after the second sparse convolution, the rectified linear unit and the first sigmoid function; The first modality information is removed from the input point cloud latent representation to obtain a third point cloud latent representation.
4. The method according to claim 3, It is characterized in that The sparse cross attention unit includes a second sparse residual block, a third sparse convolution, a feature mapping unit, a fully connected layer, and a second sigmoid function; The step of performing modal interaction processing on the point cloud features and the image features through the sparse cross attention unit to obtain a potential representation of the output point cloud comprises: Processing the point cloud features through the second sparse residual block and the third sparse convolution to obtain a fifth point cloud potential representation; Performing image mapping processing on the image features through the feature mapping unit, the fully connected layer and the second Sigmoid function to obtain a mask; Mapping the fifth point cloud potential representation to the mask to obtain second modality information; The point cloud features and the second modality information are added to obtain an output point cloud potential representation.
5. The method according to claim 1, It is characterized in that The entropy model includes a quantizer, an octree encoder, an octree decoder, a conditional checkerboard entropy model, an arithmetic encoder and an arithmetic decoder; The step of entropy encoding the first point cloud potential representation and the first view potential representation by using the entropy model to obtain a second point cloud potential representation and a second view potential representation comprises: Splitting the first point cloud potential representation into first point cloud coordinates and first point cloud features; Performing quantization operations on the first point cloud feature and the first view potential representation respectively by the quantizer to obtain a point cloud feature quantized potential representation and a view feature quantized potential representation; Compressing the first point cloud coordinates by the octree encoder and the octree decoder to obtain point cloud compressed coordinates; Performing symbol probability estimation on the point cloud feature quantization potential representation and the view feature quantization potential representation through the conditional checkerboard entropy model to obtain a view entropy parameter, a point cloud anchor point entropy parameter, and a point cloud non-anchor point entropy parameter; Through the arithmetic encoder and the arithmetic decoder, a second point cloud potential representation and a second view potential representation are obtained based on the point cloud feature quantization potential representation, the view feature quantization potential representation, the point cloud compressed coordinates, the view entropy parameters, the point cloud anchor point entropy parameters and the point cloud non-anchor point entropy parameters.
6. The method according to claim 5, It is characterized in that The conditional checkerboard entropy model includes a super arithmetic encoder, a super arithmetic decoder, a super encoder, a super decoder, a potential transformation module, a potential recovery module, a context prediction module and an entropy parameter prediction module; The step of performing symbol probability estimation on the point cloud feature quantization potential representation and the view feature quantization potential representation through the conditional checkerboard entropy model to obtain view entropy parameters, point cloud anchor point entropy parameters and point cloud non-anchor point entropy parameters comprises: The point cloud feature quantization potential representation and the view feature quantization potential representation are processed by the super encoder, the super arithmetic encoder, the super arithmetic decoder and the super decoder respectively to obtain point cloud space distribution information and view space distribution information; Processing the view space distribution information by the entropy parameter prediction module to obtain a view entropy parameter; Splitting the point cloud spatial distribution information and the view spatial distribution information into anchor point potential representations and non-anchor point potential representations respectively; Performing modal fusion on the anchor point potential representation and the non-anchor point potential representation through the potential transformation module to obtain an anchor point potential representation and a non-anchor point potential representation; Processing the anchor point potential representation and the anchor point potential representation of the point cloud spatial distribution information by the entropy parameter prediction module to obtain a point cloud anchor point entropy parameter; The non-anchor potential representation and the context anchor potential representation of the point cloud spatial distribution information are processed by the entropy parameter prediction module to obtain point cloud non-anchor entropy parameters, wherein the context anchor potential representation is obtained by processing the point cloud anchor entropy parameters by the potential recovery module and the context prediction module to obtain the context, and connecting the context and the non-anchor potential representation.
7. The method according to claim 1, It is characterized in that The decoder comprises a second combination module, a residual feature upsampling module, a quality enhancement module and a second modality redundancy removal module, wherein the second combination module comprises an upsampling unit, a residual network unit and a sparse self-attention unit; The step of decoding the second point cloud potential representation and the second view potential representation by the decoder to obtain a reconstructed point cloud comprises: Processing the second point cloud potential representation by the second combination module to obtain an output point cloud; Processing the second view potential representation by the residual feature upsampling module to obtain an output view, and performing modal redundancy removal processing on the output point cloud and the output view by the second modal redundancy removal module; The quality enhancement module performs quality enhancement processing on the output point cloud and the output view to obtain a reconstructed point cloud.
8. A point cloud compression architecture, It is characterized in that The architecture includes an encoder, an entropy model, and a decoder; The encoder is used to perform joint encoding processing on the original point cloud and the corresponding original view to obtain a first point cloud potential representation and a first view potential representation; The entropy model is used to perform entropy encoding on the first point cloud potential representation and the first view potential representation to obtain a second point cloud potential representation and a second view potential representation; The decoder is used to decode the second point cloud potential representation and the second view potential representation to obtain a reconstructed point cloud.
9. An electronic device, It is characterized in that The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the point cloud compression method according to any one of claims 1 to 7.
10. A storage medium, It is characterized in that The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the point cloud compression method according to any one of claims 1 to 7 are implemented.
