Dynamic point cloud compression method and system based on implicit modeling and detail reconstruction
By employing implicit modeling and detail reconstruction methods, the problems of inaccurate motion modeling and loss of high-frequency details in dynamic point cloud compression are solved, achieving efficient point cloud reconstruction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2026-01-23
- Publication Date
- 2026-05-05
AI Technical Summary
Existing dynamic point cloud compression techniques rely on explicit motion vectors, leading to inaccurate motion modeling, high bitrate overhead, and loss of high-frequency geometric details during reconstruction.
The method employs implicit modeling and detail reconstruction. The FMT module implicitly aligns the features of the current frame with those of historical reference frames, and the CTR module adaptively aggregates local information to generate high-quality point cloud reconstruction.
It eliminates the need for explicit motion vector encoding, reduces bit rate overhead, significantly improves the visual quality and geometric fidelity of reconstructed point clouds, and restores high-frequency details.
Smart Images
Figure CN121985143A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and 3D graphics technology, and in particular to a dynamic point cloud compression method and system based on implicit modeling and detail reconstruction. Background Technology
[0002] Dynamic point cloud (DPC) is a temporal data series composed of a sparsely and unordered distribution of point cloud frames in three-dimensional space. As an emerging form of three-dimensional content representation, it has shown broad application prospects in fields such as immersive media, motion capture, autonomous driving, and AR / VR. However, its massive data volume poses a severe challenge to storage and real-time transmission, making efficient compression technology a key prerequisite for practical deployment. Current dynamic point cloud compression methods mainly eliminate temporal redundancy through inter-frame prediction. Its core lies in using reference frame information to predict the current frame, encoding only the prediction residual. Mainstream technologies can be divided into two categories: standardized schemes and deep learning-driven schemes, but both have significant limitations, as follows: 1. Standardized methods rely on explicit motion vector coding. G-PCC and V-PCC, defined by MPEG, are currently widely used compression standards. G-PCC uses an octree structure to predict and encode geometry, while V-PCC achieves compression through 2D projection combined with a video encoder (such as HEVC). Both require explicit estimation and encoding of motion vectors (MVs) to indicate the point correspondence from the reference frame to the current frame. 2. Based on the "Motion+Residual" paradigm of neural compression methods, recent deep learning-based dynamic point cloud compression (DPCC) methods (such as D-DPCC and AdaDPCC) have introduced variational autoencoders (VAEs) to compress spatial redundancy, but they still heavily rely on explicit motion modeling in the temporal dimension. A typical approach is to use KNN or 3DAWI (3D Adaptively Weighted Interpolation) algorithms to estimate motion vectors in the latent feature space, and then entropy-encode the prediction residuals.
[0003] Existing dynamic point cloud compression technologies suffer from fundamental bottlenecks in motion modeling and detail fidelity: on the one hand, explicit motion vectors struggle to adapt to the irregular dynamics of point clouds; on the other hand, even with implicit modeling, there is a lack of effective cross-frame detail recovery mechanisms to reconstruct high-frequency geometric structures. Therefore, a novel compression method is urgently needed that eliminates the need for explicit motion vectors and adaptively fuses cross-frame local information at the decoding end to recover fine details, achieving a balance between high efficiency and high quality. Summary of the Invention
[0004] The purpose of this invention is to provide a dynamic point cloud compression method and system based on implicit modeling and detail reconstruction, which solves the problems of inaccurate motion modeling, high bit rate overhead, and loss of high-frequency geometric details during the restoration and reconstruction process caused by the reliance on explicit motion vectors in existing dynamic point cloud compression.
[0005] To achieve the above objectives, this invention provides a dynamic point cloud compression method based on implicit modeling and detail reconstruction, comprising the following steps: S1, Obtain the dynamic point cloud sequence Each frame at time t , In three-dimensional coordinates, d represents the voxel occupancy feature, N is the number of point clouds, and d is the feature dimension. S2, for the current frame point cloud Multi-scale downsampling is performed to obtain latent representations at k scales: ,in , Let k be the potential representation at the k-th scale. Let k be the coordinates at the k-th scale. For features at the k-th scale, k is not less than 3; S3, the coordinates of the lowest resolution in the current frame. Reconstructed historical reference frames stored in the reference buffer The input FMT module implicitly aligns features from historical reference frames to the current frame, outputting contextual features with motion awareness capabilities. ; S4, the feature with the lowest resolution in the current frame. and After concatenation, the data is encoded by a context encoder to generate a higher-level latent representation. ; S5 uses a differentiated compression strategy to process the bitstream and outputs the compressed bitstream. S6 receives the compressed bitstream and decodes it. , , Constructing high-level potential representations ; S7 will restore the decoded data. and the reconstructed historical reference frames stored in the reference buffer Input the FMT module to rebuild the context representation. = ; S8, will, In , In Input context decoder fusion to generate temporally aligned features ; S9, a feature for timing alignment The input CTR module adaptively aggregates local information from historical reference frames to obtain optimized features. ,Will and Combined, a potential point cloud is generated. ; S10, Full-resolution reconstructed point cloud generated through multi-scale upsampling .
[0006] Preferably, in S3 and S7, the implementation steps of the FMT module include: S301, in the lowest resolution coordinate space of the current frame. In the middle, construct the current frame With historical reference frames A fixed adjacency matrix between them; S302, based on adjacency relationships, extract the relative position information between adjacent points and the corresponding reference frame features, and aggregate them; S303, input the aggregated features into the Softmax function to generate a soft mask with a value range of [0,1]; S304, Soft mask is used to mask neighborhood features in the reference frame. Weighted modulation is performed to obtain weighted context features, which are then compared with the anchor coordinates of the current frame. The modulated features are concatenated along the channel dimension and remapped through a lightweight multilayer perceptron to finally output the motion-aligned feature representation of the current frame.
[0007] Preferably, in S301, the adjacency matrix includes a KNN adjacency matrix, a ball query adjacency matrix, a voxel nearest neighbor matrix, and a learned graph structure matrix.
[0008] Preferably, in S9, the implementation steps of the CTR module include: S901, query features derived from the aligned point cloud at time t. Key features and value features are derived from historical reference point clouds. ; S902, query features, key features, and value features are specifically represented by the following formulas: ; ; ; In the formula, i is the index of the i-th point in the current point cloud; j is the index of the j-th point in the reference point cloud; The query feature for the i-th point; The key feature of the j-th point; The value characteristic of the j-th point; To query the feature transformation function; For key feature transformation function; For value-based characteristic transformation functions; Let i be the i-th point in the rearrangement; The j-th point in the reference point cloud; S903, for each query point i, first determine the local neighborhood in its reference point cloud, and then calculate the attention output, as shown in the following formula: ; In the formula, It is a learnable position encoding function used to characterize the geometric relationship between the query point and the reference point; For a multilayer perceptron, output a vector attention weight for each feature channel; ⊙ represents the Softmax normalization operation at neighborhood points; ⊙ represents channel-level multiplication, which modulates the value features channel by channel. Let i be the coordinates of the i-th point in the point cloud refined in the k-th stage; This is the k-th refinement stage; Let j be the coordinates of the reference point cloud in the k-th stage; The k-th stage of sampling for the reference frame; Let i be the i-th point in the point cloud; Let i be the i-th point in the k-th permutation of the point cloud; The j-th point in the k-th stage of the reference point cloud; S904, the aggregated features are fused with the original features through residual connections and layer normalization to obtain a refined feature representation.
[0009] Preferably, in S902, the specific content is implemented through linear projection or a lightweight multilayer perceptron.
[0010] Preferably, in S903, the attention includes cross attention, scalar attention, linear attention, and lightweight attention.
[0011] Preferably, in S5, the compression strategy includes: feature Compression is achieved using a conditional entropy model; coordinate Compression is performed using an octree-based G-PCC encoder; coordinate Compression is achieved through an end-to-end learnable lossless codec.
[0012] Preferably, in S6, After compression, it is then decoded using a conditional entropy model. After compression, it is then decoded using the G-PCC decoder. The compressed data is then decoded using a lossless decoder.
[0013] A dynamic point cloud compression system based on implicit modeling and detail reconstruction includes: FMT module: used to receive the lowest resolution coordinates and historical reference frames, implicitly align the historical reference frame features, and output motion-aware context features; CTR module: Used to adaptively aggregate local information from historical reference frames and output optimized features; Downsampling module: used to perform multi-scale downsampling on the point cloud of the current frame to generate a multi-scale latent representation; Upsampling module: used to perform multi-scale upsampling on potential point clouds to generate full-resolution reconstructed point clouds; Context encoder: used to concatenate the lowest resolution features of the current frame with motion-aware context features to encode and generate higher-level latent representations; Context decoder: used to fuse high-level latent features with contextual representations to generate temporally aligned features; Reference buffer: Used to store reconstructed point clouds and provide historical reference frames for subsequent frames.
[0014] Preferably, the upsampling module and the downsampling module are constructed by sparse convolution.
[0015] Therefore, the present invention employs the above-mentioned dynamic point cloud compression method and system based on implicit modeling and detail reconstruction, which has the following beneficial effects: 1. No explicit motion vector encoding required, reducing bit rate overhead: The FMT module implicitly models the dynamic correspondence between the current frame and historical reference frames in the latent feature space through a fixed adjacency structure and learnable soft mask. This eliminates the need for explicit estimation, encoding, and transmission of motion vectors, significantly reducing bit rate overhead and improving the robustness of modeling non-rigid motions such as complex deformations and occlusions.
[0016] 2. Accurate restoration of cross-frame details, improving reconstruction quality: The decoding end introduces a CTR module, which adaptively aggregates local geometric and semantic information from multiple historical reference frames through a vector attention mechanism. This allows for the fine-grained repair of high-frequency details (such as edges, corners, and microstructures) lost after motion alignment, effectively eliminating local alignment artifacts and significantly improving the visual quality and geometric fidelity of the reconstructed point cloud.
[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0018] Figure 1This is a diagram of the end-to-end dynamic point cloud coding framework according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the FMT module according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the CTR module according to an embodiment of the present invention; Figure 4 This is a detailed schematic diagram (sparse convolution) of the upsampling module and the downsampling module in an embodiment of the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments in this application without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.
[0020] It should be noted that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products, or devices.
[0021] Similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0022] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed when in use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.
[0023] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," and "connect" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0024] Example like Figure 1 As shown, this embodiment provides a dynamic point cloud compression method based on implicit modeling and detail reconstruction, including the following steps: k takes the value of 3, and n takes the value of 1.
[0025] S1, Obtain the dynamic point cloud sequence Each frame at time t , In three-dimensional coordinates, d represents the voxel occupancy feature, N is the number of point clouds, and d is the feature dimension.
[0026] S2, for the current frame point cloud Multi-scale downsampling is performed, sequentially passing the data through three downsampling modules to obtain latent representations at three scales: ,in , This is a potential representation of the third scale. The coordinates are for the third scale. This is a feature of the third scale.
[0027] S3, the coordinates of the lowest resolution in the current frame. Reconstructed historical reference frames stored in the reference buffer The input FMT module implicitly aligns features from historical reference frames to the current frame, outputting contextual features with motion awareness capabilities. .
[0028] The FMT module uses the reconstructed historical reference frame to perform a forward feature reference on the current frame.
[0029] like Figure 2 As shown, the implementation steps of the FMT module include: S301, in the lowest resolution coordinate space of the current frame. In the middle, construct the current frame With historical reference frames A fixed KNN adjacency matrix between them.
[0030] S302, based on adjacency relationships, extracts the relative position information between adjacent points and the corresponding reference frame features, and aggregates them.
[0031] S303, input the aggregated features into the Softmax function to generate a learnable, dynamically changing soft mask with a value range of [0,1].
[0032] S304, Soft mask is used to mask neighborhood features in the reference frame. Weighted modulation is performed to obtain weighted context features, which are then compared with the anchor coordinates of the current frame. The modulated features are concatenated along the channel dimension and remapped through a lightweight multilayer perceptron (MLP) to finally output the motion-aligned feature representation of the current frame.
[0033] This invention employs a fixed KNN adjacency matrix (i.e., the adjacency structure is not updated during training and inference), which significantly improves the stability and convergence of model training compared to methods that dynamically update adjacency relationships in each iteration.
[0034] In S301, the adjacency matrix includes the KNN adjacency matrix, the ball query adjacency matrix, the voxel nearest neighbor matrix, and the learning graph structure matrix.
[0035] S4, the feature with the lowest resolution in the current frame. and After concatenation, the data is encoded by a context encoder to generate a higher-level latent representation. .
[0036] S5 uses a differentiated compression strategy to process the bitstream and outputs the compressed bitstream.
[0037] Compression strategies include: feature Compression is achieved using a conditional entropy model.
[0038] coordinate Compression is performed using an octree-based G-PCC encoder.
[0039] coordinate Compression is achieved through an end-to-end learnable lossless codec to preserve the structural details required for decoding end-alignment.
[0040] S6 receives the compressed bitstream. After compression, it is then decoded using a conditional entropy model. After compression, it is then decoded using the G-PCC decoder. The compressed data is then decoded using a lossless decoder to restore the original data. , , Constructing high-level potential representations .
[0041] S7 will restore the decoded data. and the reconstructed historical reference frames stored in the reference buffer Input the FMT module to rebuild the context representation. = .
[0042] S8, will, In , In Input context decoder fusion to generate temporally aligned features .
[0043] S9, a feature for timing alignment The input CTR module adaptively aggregates local information from historical reference frames, effectively recovering high-frequency details lost due to motion transformation or quantization, and obtaining optimized features. ,Will and Combined, a potential point cloud is generated. .
[0044] In the feature alignment stage of the decoder, this invention employs an FMT module to achieve global feature alignment between the current frame and historical reference frames. Although FMT can effectively reconstruct the global structure, significant alignment artifacts still exist in local areas, manifesting as geometric deformation and detail misalignment. To alleviate such local misalignment problems, this invention introduces a point-based CTR module in the decoder for fine-grained feature optimization, thereby enhancing the geometric consistency of point cloud features and the accuracy of local detail alignment.
[0045] Inspired by self-attention mechanisms, attention can be categorized into scalar and vector types. Vector attention is more suitable for 3D point clouds because it can capture local geometric changes through directionality and inter-channel feature dependencies. To establish fine-grained correspondences between two frames of point clouds, this invention extends vector attention from self-attention to a cross-attention mechanism.
[0046] like Figure 3 As shown, the implementation steps of the CTR module include: S901, query features derived from the aligned point cloud at time t. Key features and value features are derived from historical reference point clouds. .
[0047] S902, query features, key features, and value features are specifically represented by the following formulas: ; ; ; In the formula, i is the index of the i-th point in the current point cloud; j is the index of the j-th point in the reference point cloud; The query feature for the i-th point; The key feature of the j-th point; The value characteristic of the j-th point; To query the feature transformation function; For key feature transformation function; For value-based characteristic transformation functions; Let i be the i-th point in the rearrangement; The j-th point in the reference point cloud; S903, for each query point i, first determine the local neighborhood in its reference point cloud, and then calculate the cross-attention output, as shown in the following formula: ; In the formula, It is a learnable position encoding function used to characterize the geometric relationship between the query point and the reference point; For a multilayer perceptron, output a vector attention weight for each feature channel; ⊙ represents the Softmax normalization operation at neighborhood points; ⊙ represents channel-level multiplication, which modulates the value features channel by channel. Let i be the coordinates of the i-th point in the refined point cloud of the third stage. This is the third stage of refinement; The coordinates of the j-th point in the reference point cloud for stage 3; The third stage of sampling for the reference frame; Let i be the i-th point in the point cloud; This refers to the i-th point in the third permutation of the point cloud; The j-th point in the third stage of the reference point cloud; S904, the aggregated features are fused with the original features through residual connections and layer normalization to obtain a refined feature representation.
[0048] In S902, the specific content is implemented through linear projection or a lightweight multilayer perceptron.
[0049] In S903, attention includes cross attention, scalar attention, linear attention, and lightweight attention.
[0050] S10, Multi-scale upsampling is performed, and the data is processed through three upsampling modules to generate a full-resolution reconstructed point cloud. It is stored in the reference buffer for prediction of subsequent frames.
[0051] A dynamic point cloud compression system based on implicit modeling and detail reconstruction includes: FMT module: Used to receive the lowest resolution coordinates and historical reference frames, implicitly align the historical reference frame features, and output motion-aware context features.
[0052] CTR module: Used to adaptively aggregate local information from historical reference frames and output optimized features.
[0053] Downsampling module: Used to perform multi-scale downsampling on the point cloud of the current frame to generate a multi-scale latent representation.
[0054] Upsampling module: Used to perform multi-scale upsampling on potential point clouds to generate full-resolution reconstructed point clouds.
[0055] Context encoder: Used to concatenate the lowest resolution features of the current frame with motion-aware context features to encode and generate higher-level latent representations.
[0056] Context decoder: used to fuse high-level latent features with contextual representations to generate temporally aligned features.
[0057] Reference buffer: Used to store reconstructed point clouds and provide historical reference frames for subsequent frames.
[0058] like Figure 4 As shown, the upsampling module and the downsampling module are constructed through sparse convolution.
[0059] Therefore, the present invention employs the above-mentioned dynamic point cloud compression method and system based on implicit modeling and detail reconstruction, which can solve the problems of inaccurate motion modeling, high bit rate overhead, and loss of high-frequency geometric details during the restoration and reconstruction process caused by the reliance on explicit motion vectors in existing dynamic point cloud compression.
[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A dynamic point cloud compression method based on implicit modeling and detail reconstruction, characterized in that: Includes the following steps: S1, Obtain the dynamic point cloud sequence Each frame at time t , In three-dimensional coordinates, d represents the voxel occupancy feature, N is the number of point clouds, and d is the feature dimension. S2, for the current frame point cloud Multi-scale downsampling is performed to obtain latent representations at k scales: ,in , Let k be the potential representation at the k-th scale. Let k be the coordinates at the k-th scale. For features at the k-th scale, k is not less than 3; S3, the coordinates of the lowest resolution in the current frame. Reconstructed historical reference frames stored in the reference buffer The input FMT module implicitly aligns features from historical reference frames to the current frame, outputting contextual features with motion awareness capabilities. ; S4, the feature with the lowest resolution in the current frame. and After concatenation, the data is encoded by a context encoder to generate a higher-level latent representation. ; S5 uses a differentiated compression strategy to process the bitstream and outputs the compressed bitstream. S6 receives the compressed bitstream and decodes it. , , Constructing high-level potential representations ; S7 will restore the decoded data. and the reconstructed historical reference frames stored in the reference buffer Input the FMT module to rebuild the context representation. = ; S8, In , In Input context decoder fusion to generate temporally aligned features ; S9, a feature for timing alignment The input CTR module adaptively aggregates local information from historical reference frames to obtain optimized features. ,Will and Combined, a potential point cloud is generated. ; S10, Full-resolution reconstructed point cloud generated through multi-scale upsampling .
2. The dynamic point cloud compression method based on implicit modeling and detail reconstruction according to claim 1, characterized in that: In S3 and S7, the implementation steps of the FMT module include: S301, in the lowest resolution coordinate space of the current frame. In the middle, construct the current frame With historical reference frames The fixed adjacency matrix between them; S302, based on adjacency relationships, extract the relative position information between adjacent points and the corresponding reference frame features, and aggregate them; S303, input the aggregated features into the Softmax function to generate a soft mask with a value range of [0,1]; S304, Soft mask is used to mask neighborhood features in the reference frame. Weighted modulation is performed to obtain weighted context features, which are then compared with the anchor coordinates of the current frame. The modulated features are concatenated along the channel dimension and remapped through a lightweight multilayer perceptron to finally output the motion-aligned feature representation of the current frame.
3. The dynamic point cloud compression method based on implicit modeling and detail reconstruction according to claim 2, characterized in that: In S301, the adjacency matrix includes the KNN adjacency matrix, the ball query adjacency matrix, the voxel nearest neighbor matrix, and the learning graph structure matrix.
4. The dynamic point cloud compression method based on implicit modeling and detail reconstruction according to claim 1, characterized in that: In S9, the implementation steps of the CTR module include: S901, query features derived from the aligned point cloud at time t. Key features and value features are derived from historical reference point clouds. ; S902, query features, key features, and value features are specifically represented by the following formulas: ； ； ； In the formula, i is the index of the i-th point in the current point cloud; j is the index of the j-th point in the reference point cloud; The query feature for the i-th point; The key feature of the j-th point; The value characteristic of the j-th point; To query the feature transformation function; For key feature transformation function; For value-based characteristic transformation functions; Let i be the i-th point in the rearrangement; The j-th point in the reference point cloud; S903, for each query point i, first determine the local neighborhood in its reference point cloud, and then calculate the attention output, as shown in the following formula: ； In the formula, It is a learnable position encoding function used to characterize the geometric relationship between the query point and the reference point; For a multilayer perceptron, output a vector attention weight for each feature channel; ⊙ represents the Softmax normalization operation at neighborhood points; ⊙ represents channel-level multiplication, which modulates the value features channel by channel. Let i be the coordinates of the i-th point in the point cloud refined in the k-th stage; This is the k-th refinement stage; Let j be the coordinates of the reference point cloud in the k-th stage; The k-th stage of sampling for the reference frame; Let i be the i-th point in the point cloud; Let i be the i-th point in the k-th permutation of the point cloud; The j-th point in the k-th stage of the reference point cloud; S904, the aggregated features are fused with the original features through residual connections and layer normalization to obtain a refined feature representation.
5. The dynamic point cloud compression method based on implicit modeling and detail reconstruction according to claim 4, characterized in that: In S902, the specific content is implemented through linear projection or a lightweight multilayer perceptron.
6. The dynamic point cloud compression method based on implicit modeling and detail reconstruction according to claim 4, characterized in that: In S903, attention includes cross attention, scalar attention, linear attention, and lightweight attention.
7. The dynamic point cloud compression method based on implicit modeling and detail reconstruction according to claim 1, characterized in that: In S5, compression strategies include: feature Compression is achieved using a conditional entropy model; coordinate Compression is performed using an octree-based G-PCC encoder; coordinate Compression is achieved through an end-to-end learnable lossless codec.
8. The dynamic point cloud compression method based on implicit modeling and detail reconstruction according to claim 1, characterized in that: In S6, After compression, it is then decoded using a conditional entropy model. After compression, it is then decoded using the G-PCC decoder. The compressed data is then decoded using a lossless decoder.
9. A dynamic point cloud compression system based on implicit modeling and detail reconstruction, employing the dynamic point cloud compression method based on implicit modeling and detail reconstruction as described in any one of claims 1-8, characterized in that: include: FMT module: used to receive the lowest resolution coordinates and historical reference frames, implicitly align the historical reference frame features, and output motion-aware context features; CTR module: Used to adaptively aggregate local information from historical reference frames and output optimized features; Downsampling module: used to perform multi-scale downsampling on the point cloud of the current frame to generate a multi-scale latent representation; Upsampling module: used to perform multi-scale upsampling on potential point clouds to generate full-resolution reconstructed point clouds; Context encoder: used to concatenate the lowest resolution features of the current frame with motion-aware context features to encode and generate higher-level latent representations; Context decoder: used to fuse high-level latent features with contextual representations to generate temporally aligned features; Reference buffer: Used to store reconstructed point clouds and provide historical reference frames for subsequent frames.
10. A dynamic point cloud compression system based on implicit modeling and detail reconstruction according to claim 9, characterized in that: The upsampling module and the downsampling module are constructed using sparse convolution.