Attention unmasking dual-branch distillation method for three-dimensional point cloud completion

Abstract

The application discloses a kind of attention anti-mask double-branch distillation methods for three-dimensional point cloud completion.The method first extracts multi-scale spatial attention map from teacher model, and generates anti-mask point cloud based on it;Then build double-branch student network, respectively process the intermediate feature point cloud of student model and the anti-mask point cloud;Then the joint recursive distillation module is used to the point cloud features of two student branches and teacher model Recursive upsampling and feature aggregation under geometric consistency constraint;Finally, using triple supervision optimization module, jointly optimize CD distance loss, feature distillation loss and cross-branch contrast loss to train student model.The application constructs teacher attention guided anti-mask mechanism, introduces double-input branch architecture, designs joint recursive distillation mechanism, and simultaneously constructs triple distillation loss function, effectively improves the point cloud completion precision and robustness of student model in complex autonomous driving scene.

Description

Technical Field

[0001] This invention relates to the field of 3D point cloud completion technology, and particularly to an attention-based inverse mask bibranch distillation method for 3D point cloud completion. Background Technology

[0002] With the rapid development of autonomous driving technology, environmental perception systems are placing higher demands on the integrity and reliability of data. Point cloud completion, as a fundamental task in 3D computer vision, is responsible for restoring and reconstructing incomplete, sparse, or noisy point cloud data acquired by LiDAR. Its quality directly affects the accuracy of subsequent key tasks such as semantic segmentation, target detection, and tracking.

[0003] In practical autonomous driving systems, lightweight point cloud completion models are typically deployed due to limitations in computing resources, power consumption, and real-time performance of in-vehicle embedded platforms. However, these lightweight models exhibit a significant performance gap compared to teacher models with large parameter sets and complex structures, making it difficult to maintain high-precision completion results in complex scenarios. To alleviate this problem, researchers have proposed various model compression and performance enhancement techniques, including network pruning, quantization, lightweight network architecture design, and knowledge distillation. Among these, knowledge distillation, by transferring the rich knowledge contained in the teacher model to the student model, has become an effective means of improving the performance of lightweight models.

[0004] However, existing knowledge distillation methods for point cloud completion tasks still face many challenges in autonomous driving scenarios:

[0005] Low efficiency in transferring key attention regions: When dealing with sparse, occluded, or noisy point clouds, the teacher model can focus on key geometric regions (such as edges and structural details) through its internal spatial attention mechanism. Traditional distillation methods lack explicit mechanisms to guide the student model to focus on these regions, resulting in geometric distortion and loss of detail in the student model in complex scenes.

[0006] Single-path training lacks generalization ability: Existing methods usually only perform distillation training based on the original point cloud. The student model lacks the ability to adapt to real-world interference (such as noise, occlusion, and uneven density), resulting in poor performance in real-world scenarios.

[0007] Lack of multi-view feature alignment mechanism: Some methods attempt to introduce data augmentation or feature masks to improve robustness, but lack effective cross-view alignment mechanisms. Figure 1 Consistency constraints lead to conflicts in the features learned by the student model under different input conditions, affecting the structural integrity of the reconstructed point cloud.

[0008] Weak geometric consistency constraints: Existing distillation strategies are unable to convey the teacher model’s ability to understand geometric structures and model hierarchical contexts, resulting in student models having difficulty maintaining consistency between local details and global structure, which limits the accuracy of reproducing complex shapes.

[0009] In summary, although knowledge distillation has shown potential in model compression and performance improvement, existing methods still have significant shortcomings in attention guidance, multi-path training, feature alignment, and geometric consistency modeling, which limits its further application in autonomous driving point cloud completion tasks. Therefore, there is an urgent need for a knowledge distillation method that can effectively transfer the teacher model's attention mechanism and enhance the student model's generalization and structural awareness capabilities. Summary of the Invention

[0010] To address the above problems, this invention provides an attention-based inverse mask bi-branch distillation method for 3D point cloud completion. The method aims to improve the accuracy and robustness of student models in complex autonomous driving scenarios by constructing a teacher attention-guided inverse mask mechanism, introducing a dual-input branch architecture, designing a joint recursive distillation mechanism, and constructing a triple distillation loss function.

[0011] This invention provides an attention-based inverse mask bibranch distillation method for 3D point cloud completion, comprising:

[0012] S1, extract a multi-scale spatial attention map from the teacher model, and generate an inverse mask point cloud based on the attention map;

[0013] S2, construct a two-branch student network, wherein the first branch processes the intermediate feature point cloud of the student model, and the second branch processes the inverse mask point cloud;

[0014] S3, using a joint recursive distillation module to perform recursive upsampling and feature aggregation on the feature point clouds of the first branch, the second branch and the teacher model under geometric consistency constraints;

[0015] S4 utilizes a triple-supervised optimization module to jointly optimize the CD distance loss, feature distillation loss, and cross-branch contrast loss to train the student model;

[0016] S5 inputs the missing 3D point cloud into the trained student model for inference and obtains the point cloud completion result.

[0017] Furthermore, S1 includes:

[0018] S11: Extract multi-scale feature maps from the shallow, middle and deep layers of the teacher model, calculate the L2 norm of each feature map in the channel dimension, and generate multi-level spatial attention maps.

[0019] S12, the multi-level spatial attention map is upsampled to a uniform size and then weighted and fused according to the importance of the hierarchy to obtain a fused attention map, which is then normalized.

[0020] S13, Based on the normalized fused attention map, a binary mask matrix is ​​generated by applying dynamic inverse masking rules;

[0021] S14, the binary mask matrix is ​​copied along the channel dimension and multiplied element-wise with the intermediate feature point cloud learned by the student model to generate the inverse mask point cloud.

[0022] By extracting and fusing multi-scale spatial attention maps from the shallow, medium, and deep layers of the teacher model, this method can comprehensively capture key geometric region information from local details to global semantics. This multi-level attention mechanism generates more accurate and robust fused attention maps, laying a reliable foundation for the subsequent generation of high-quality inverse mask point clouds and effectively avoiding information loss caused by single-scale attention bias.

[0023] Furthermore, the dynamic inverse masking rule is specifically as follows:

[0024] For each position in the normalized fused attention map, if its normalized attention value is lower than a preset threshold, the position information is retained.

[0025] If its normalized attention value is higher than or equal to the preset threshold, the location information is retained or discarded based on random probability.

[0026] The applied dynamic inverse masking rule forces the student model to focus on low-response regions (such as object edges and occluded parts) that the teacher model ignores, while a random retention mechanism prevents the complete loss of key high-response information. This strategy significantly enhances the student model's feature completion and reasoning capabilities under conditions of missing information, thereby improving its generalization and robustness in complex real-world scenarios.

[0027] Furthermore, the specific structure of the dual-branch student network in S2 is as follows:

[0028] The first branch is the standard distillation path, whose input is the intermediate feature point cloud extracted from the original point cloud by the student model backbone network;

[0029] The second branch is the inverse mask distillation path, whose input is the inverse mask point cloud, and the path contains a simple convolution generator for mapping the features of the inverse mask point cloud to the complete semantic space of the teacher model.

[0030] The constructed dual-branch student network achieves collaborative learning and complementary capabilities. The standard distillation path ensures that the student masters the teacher model's basic representational ability under ideal input, while the inverse mask distillation path specifically trains its semantic recovery ability under incomplete features through a feature mapper. The two paths complement each other, jointly improving the overall performance of the student model, enabling it to maintain high accuracy while coping with incomplete input and noise interference.

[0031] Furthermore, the joint recursive distillation module in S3 adopts a recursive structure, performing feature upsampling and aggregation through multi-scale point cloud sets and Channel-aware Transformer operations.

[0032] The joint recursive distillation module decomposes the point cloud completion task into a recursive process from coarse to fine by using multi-scale point cloud sets and CT operations. This structured approach introduces the geometric prior knowledge of the teacher model as supervision at each scale, thereby greatly ensuring the geometric consistency and structural rationality of the final generated point cloud and avoiding detail blurring and overall deformation.

[0033] Furthermore, S3 specifically includes:

[0034] S31, the input point cloud is downsampled in layers using the farthest point sampling algorithm to obtain a point cloud set containing m scales;

[0035] S32, at the (m-1)th scale, perform a Channel-aware Transformer operation to aggregate support information from the shape point cloud and trajectory point cloud, and output the enhanced point features;

[0036] S33, the output features at the (m-1)th scale are concatenated with the query features at the (m-2)th scale, and after fusion by a multilayer perceptron, a value vector for CT operation at the (m-2)th scale is generated;

[0037] S34, continue to perform CT operation at the m-2 scale, iterate the process until high-level point cloud features are obtained.

[0038] This module achieves cross-path information aggregation and contextual modeling by dynamically acquiring shape and trajectory point clouds from the features of each branch point cloud. An explicit recursive process (from m-1 scale to higher-level features) ensures the continuity and stability of feature aggregation. This design enables the student model to progressively and meticulously recover the point cloud structure, significantly improving the accuracy of the upsampling process and the richness of detail in the generated point cloud.

[0039] Furthermore, the specific process of the Channel-aware Transformer operation includes:

[0040] Based on the query point, key point, and value point, the attention weight is calculated using a vector attention mechanism;

[0041] Based on the attention weights, multiple attention heads are used to aggregate local features to generate upsampled point features;

[0042] After concatenating the upsampled point features with the query features, weighted fusion is performed through local average pooling and weight calculation, and residual connections are added to update the upsampled point features.

[0043] The vector attention, multi-head aggregation, and channel-aware weighting mechanisms introduced in the Channel-aware Transformer operation together constitute a powerful feature upsampling engine. Vector attention can accurately capture local geometric relationships; the multi-head mechanism supports parallel learning of multiple geometric patterns; and channel-aware weighting can intelligently filter and fuse features. The combination of these three ultimately outputs high-quality upsampled point features.

[0044] Furthermore, the loss function used by the triple-supervised optimization module in S4 includes:

[0045] CD distance loss is used to directly optimize the geometric error between the student model's completed output and the real point cloud;

[0046] Feature distillation loss is used to minimize the difference between the output features of the student model's first and second branches and the teacher model's output features;

[0047] Cross-branch contrastive loss is used to constrain the consistency between the first and second branches of the student model in the output feature space.

[0048] The triple-supervised optimization module provides multi-faceted and refined guidance to the student model by jointly optimizing three complementary loss functions. The CD loss ensures geometric accuracy, the feature distillation loss guarantees knowledge transfer efficiency, and the cross-branch contrastive loss effectively eliminates feature conflicts in dual-path learning. This collaborative optimization strategy ensures balanced model performance across multiple tasks, ultimately resulting in a high-performance, robust, and lightweight point cloud completion model.

[0049] Compared with existing technologies, the beneficial effects of this invention are as follows: By extracting the multi-scale attention map of the teacher model and constructing an inverse mask point cloud, the invention innovatively guides the student model to focus on and complete difficult regions (such as edges and occluded parts) that the teacher has overlooked. This effectively solves the problem of inefficient transfer in key regions in traditional distillation methods, significantly improving the geometric accuracy and detail restoration of the completion results. The designed dual-branch network architecture, through the collaborative training of the standard path and the inverse mask path, enables the student model to simultaneously master the completion ability under both ideal and incomplete inputs, significantly improving its adaptability and generalization ability in real-world scenarios such as noise and occlusion. A joint recursive distillation module is adopted, using multi-scale recursive upsampling and layer-by-layer distillation supervision to integrate geometric consistency constraints throughout the entire feature learning process, ensuring that the generated point cloud is consistent in both local details and global structure, effectively avoiding shape distortion and surface breaks. A triple-supervised optimization module collaboratively optimizes the CD loss, feature distillation loss, and cross-branch contrast loss, simultaneously constraining model training from three dimensions: geometric accuracy, knowledge transfer, and internal consistency. Ultimately, a lightweight model that achieves a balance between accuracy, robustness, and efficiency is obtained. In summary, this invention successfully compresses and transfers the complex decision-making capabilities of large teacher models to lightweight student networks by organically integrating four major technical frameworks: attention inverse masking, bi-branch collaborative learning, joint recursive distillation, and triple-supervised optimization. The resulting model is suitable for in-vehicle embedded autonomous driving environment perception systems with stringent requirements for reliability, real-time performance, and computational resources, and has significant industrial application value. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this drawing or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this drawing. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0051] Figure 1 This is a diagram illustrating the method steps of the present invention;

[0052] Figure 2 This is an overall flowchart of the method of the present invention;

[0053] Figure 3 This is a flowchart of the inverse mask generation process of the present invention;

[0054] Figure 4 This is a diagram of the combined recursive distillation module of the present invention;

[0055] Figure 5 This is a CT structural diagram of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments provided by this invention without inventive effort are within the scope of protection of this invention.

[0057] This invention provides an attention-based inverse mask bibranch distillation method for 3D point cloud completion, such as... Figure 1 , Figure 2 As shown, the specific steps include:

[0058] S1, extract a multi-scale spatial attention map from the teacher model, and generate an inverse mask point cloud based on the attention map, such as... Figure 3 As shown, the specific steps include:

[0059] S11 extracts multi-scale feature maps from the shallow, medium and deep layers of the teacher model, calculates the L2 norm of each feature map in the channel dimension, and generates a multi-level spatial attention map.

[0060] Specifically, given an input point cloud ,in Let C represent the real number space, C represent the number of channels, and N represent the number of points. This space is input into the teacher model for learning, extracting multi-scale feature maps from different depth layers. , , These features correspond to the teacher model's perception of the local details, mid-level structure, and global semantics of the input point cloud.

[0061] For each layer of feature map Spatial attention map is generated using the L2 norm in the channel dimension. :

[0062] ;

[0063] Where C represents the number of channels.

[0064] S12, the multi-level spatial attention map is upsampled to a uniform size and then weighted and fused according to the importance of the levels to obtain a fused attention map, which is then normalized.

[0065] Specifically, attention maps at different resolutions are first upsampled to a uniform size N×K (K represents the neighborhood size) using bilinear interpolation, and then weighted and fused to generate a fused attention map A:

[0066] ;

[0067] in, Assigning weights based on hierarchical importance;

[0068] This indicates a bilinear interpolation operation.

[0069] Next, the fused attention map A is compressed to a numerical range of 0 to 1 through normalization:

[0070] ;

[0071] in, It is a constant set to prevent division by zero operations;

[0072] This is the normalized fusion attention map.

[0073] S13, based on the normalized fused attention map, a binary mask matrix is ​​generated by applying dynamic inverse masking rules.

[0074] To address the issues raised in the normalized results, this invention designs a dynamic anti-masking rule: for each position in the normalized fused attention map, if its normalized attention value is lower than a preset threshold, the position information is retained; if its normalized attention value is higher than or equal to the preset threshold, the position information is retained or discarded based on random probability. This rule forces the student model to pay attention to low-response regions ignored by the teacher, while also preventing the complete loss of key information through a random retention mechanism.

[0075] Specifically, based on the normalized fusion attention map Design a dynamic inverse masking rule: for each point's position (i, j), if its attention value... If so, then retain the location information; if Then, randomly retain or discard, ultimately obtaining a binary mask matrix. :

[0076] ;

[0077] in, It is a random matrix of the same size as the attention map. Let be a random number at (i, j) in the random matrix;

[0078] p is a random probability, used to introduce a random retention mechanism;

[0079] The inverse mask strength can be adjusted by setting a preset threshold.

[0080] S14, the binary mask matrix is ​​copied along the channel dimension and multiplied element-wise with the intermediate feature point cloud learned by the student model to generate the inverse mask point cloud.

[0081] The generated binary mask matrix is ​​copied along the channel dimension and multiplied element-wise with the intermediate output of the student model to generate a semantically guided inverse mask point cloud. This inverse mask point cloud preserves two types of regions: low-response regions of the teacher (such as object edges, sparse regions, and occlusion boundaries) and randomly sampled high-response region fragments (such as local parts of complete surfaces and structural key points), thereby guiding the student model to learn the ability to complete features under conditions of missing key information.

[0082] Specifically, the original point cloud x The student model backbone network is input for learning, and intermediate feature point clouds are extracted. :

[0083] ;

[0084] in, The backbone network for the student model.

[0085] The generated binary mask matrix is ​​copied along the channel dimension and multiplied element-wise with the intermediate feature point cloud learned by the student model to generate the inverse mask point cloud. :

[0086] ;

[0087] in, It is a binary mask matrix;

[0088] This represents element-wise product.

[0089] S2, construct a two-branch student network, where the first branch processes the intermediate feature point cloud of the student model and the second branch processes the inverse mask point cloud.

[0090] This invention proposes dividing the student model into two branches to process intermediate feature point clouds and inverse mask point clouds, respectively. The first branch is a standard distillation path, which learns the feature representation of the teacher model on the original point cloud. Its input is the intermediate feature point cloud extracted from the original point cloud by the student model's backbone network. The second branch is an inverse mask distillation path, which learns the teacher model's representation ability in low-attention regions. Its input is the inverse mask point cloud, and this path includes a simple convolution generator to map the features of the inverse mask point cloud to the teacher model's complete semantic space.

[0091] S3 employs a joint recursive distillation module to perform recursive upsampling and feature aggregation on the feature point clouds of the first branch, the second branch, and the teacher model under geometric consistency constraints.

[0092] For the first branch, the intermediate feature point cloud is input into the joint recursive distillation module to obtain the first point cloud feature. :

[0093] ;

[0094] JRD stands for Joint Recursive Distillation Module.

[0095] For the second branch, the inverse mask point cloud The data is fed into a simple convolution generator G, and then the second point cloud features are obtained through a joint recursive distillation module. :

[0096] .

[0097] For the teacher branch, the original point cloud is input into the teacher model, and then the teacher point cloud features are obtained through a joint recursive distillation model. :

[0098] ;

[0099] ;

[0100] in, The backbone network for the teacher model;

[0101] The intermediate feature point cloud output by the teacher model.

[0102] The point cloud features output by the teacher model and the student model are aligned, and L2 normalization is used to enhance the alignment effect.

[0103] ;

[0104] This structure enables the student model to both fit the overall expression of the teacher model and capture low-attention semantics that the teacher has not significantly modeled, thus forming an effective collaborative learning mechanism.

[0105] This invention proposes a "joint recursive distillation" method to achieve refined guidance and structured constraints on the feature learning process of student models. Specifically, as follows... Figure 4 As shown, this method first constructs a multi-scale point cloud set based on the farthest point sampling algorithm, and at each scale, it adopts the "parent seed joint constraint" mechanism to dynamically aggregate the topological structure information of the seed point and the displacement semantic features of the parent point to construct a cross-level shape context representation as the basis for structure guidance.

[0106] Subsequently, in the recursive refinement stage, knowledge distillation is performed layer by layer on the student model, enabling it to gradually recover the point cloud structure from coarse to fine, thus enhancing its understanding of local geometric relationships and global shape. This joint modeling and recursive distillation approach not only improves the geometric consistency and detail recovery capability of the student model, but also significantly enhances the interpretability and controllability of the entire upsampling process, thereby improving the accuracy and robustness of the point cloud completion task.

[0107] Specifically, this method introduces a recursive structure to learn the spatial relationships between points and enhances the geometric consistency of the generation process through dual constraints, thereby better predicting the shape of missing points. Given the coordinates of the query point cloud... and corresponding features The given support point cloud consists of two parts, one of which is the shape point cloud. and corresponding features (derived from seed point cloud interpolation), and secondly, through jump connections of the input trajectory point cloud. and corresponding features .

[0108] The query point cloud represents the point cloud of the input joint recursive distillation module under the current path (for the first branch of the teacher model and the student model, the input is their respective intermediate feature point clouds; for the second branch of the student model, the input is the inverse mask point cloud), and its coordinates and features are as follows: and The supported point cloud consists of two parts: one is a shape point cloud obtained by interpolation from the intermediate feature point cloud under the current path, and the other is a point cloud with coordinates and features respectively. and The second is the trajectory point cloud (same as the query point cloud), whose coordinates and features are as follows: and .

[0109] S31, the input point cloud is downsampled in layers using the farthest point sampling algorithm to obtain a point cloud set containing m scales (in this method, m is set to 3, and index 0 corresponds to the initial input point cloud).

[0110] S32, at the (m-1)th scale, performs a Channel-aware Transformer operation, aggregating support information from the shape point cloud and trajectory point cloud, and outputting enhanced point features.

[0111] Specifically, the enhanced point features are calculated using the following formula. :

[0112] ;

[0113] in, , ( (These can represent one of the supporting point clouds) are the query vector and key vector that participate in the calculation, respectively.

[0114] Then by and The cascaded result is obtained by applying an MLP.

[0115] S33, the output features at scale m-1 are concatenated with the query features at scale m-2, and after fusion by a multilayer perceptron, a value vector for CT operation at scale m-2 is generated.

[0116] Specifically, will , and After concatenation, feature fusion is performed using an MLP to obtain a new value vector. .

[0117] S34, continue performing the CT operation at the (m-2)th scale, iterating this process until high-level point cloud features are obtained. It represents a set of generated point features.

[0118] The CT operation aims to upsample the input point features, and its structure is as follows: Figure 5 As shown. In the first l At scale, point-by-point query given input ,key Sum First, calculate the vector attention weights in the neighborhood (k nearest neighbors) of each query point based on the subtraction relationship. :

[0119] ;

[0120] in, , and It is a mapping function implemented using a multilayer perceptron (MLP);

[0121] It is a positional encoding used to learn the spatial relationships of point clouds;

[0122] It is a query vector;

[0123] It is a key vector;

[0124] To generate point features, multiple attention heads are used to learn specific geometric patterns in local features, hence the definition is... , ,in Corresponding upsampling rate, indicating The process of generating point features.

[0125] Then, the weights are adjusted using the softmax function. Normalization:

[0126] .

[0127] Subsequently, local features are aggregated to generate new point features. :

[0128] ;

[0129] in, It is a mapping function;

[0130] It is a value vector.

[0131] All Grouping generates upsampling point features .

[0132] Next, With query features Cascading yields cascade features For each query point's neighborhood (constructed via KNN and then mapped using MLP), local average pooling is applied to compress the information. This is then processed using MLP and the sigmoid activation function to obtain the final weight W. Finally, the weights are updated as follows: :

[0133] ;

[0134] in, This indicates element-wise addition, used to add residual information.

[0135] S4 utilizes a triple-supervised optimization module to jointly optimize the CD distance loss, feature distillation loss, and cross-branch contrast loss to train the student model.

[0136] To achieve refined knowledge transfer and ensure the synergy of multi-branch learning, this method jointly optimizes three complementary loss functions: the CD distance loss acts on the output of the first branch, directly optimizing the basic completion accuracy of the student model; the feature distillation loss (based on the similarity measure of normalized features) acts on both branches simultaneously, forcing their high-level feature representations to be as close as possible in direction to the corresponding features of the teacher model, achieving efficient knowledge transfer; and the cross-branch contrastive loss constrains the features adapted by the second branch (inverse masking processing path) to maintain a high degree of consistency with the features of the first branch (standard processing path) in the unmasked information preservation region. This loss effectively solves the potential geometric feature conflict problem in multi-view feature learning, significantly improving the local detail integrity, surface continuity, and overall structural consistency of the student model when reconstructing point clouds, making the completion results perform better in terms of shape rationality and geometric accuracy.

[0137] Chamfer Distance Loss: The Chamfer Distance based on the L1 norm is used to directly optimize the matching between the completed output of the student model and the real point cloud.

[0138] ;

[0139] in, The CD distance loss function;

[0140] P For predicting point cloud sets;

[0141] p For predicting a point in the point cloud set;

[0142] Q A collection of real point clouds;

[0143] q It is a point in the real point cloud set.

[0144] Characteristic distillation loss: Distillation loss combined with two branches and This results in total characteristic distillation loss. .

[0145] ;

[0146] All features have undergone L2 normalization to ensure that distillation focuses on directional information rather than amplitude.

[0147] Cross-branch comparison loss: Introducing cross-branch comparison loss This constrains the consistency of the representation of the two input branches in the output space.

[0148] ;

[0149] By using contrastive loss to compensate for differences in feature space, the student network's ability to fit unlabeled regions is improved, thus enhancing its structure-aware representation.

[0150] The total loss function is derived from a weighted combination of the three types of losses mentioned above:

[0151] ;

[0152] in, , , These are dynamically weighted coefficients; by adjusting these coefficients, specific objectives can be optimized at different stages of training.

[0153] The recommended value is 1.0 to balance geometric accuracy and training stability;

[0154] The recommended value is 0.5 to avoid overly strong feature constraints that could lead to underfitting of the student model.

[0155] The recommended value is 0.1 to prevent contrastive loss from dominating the training process and ignoring geometric consistency.

[0156] S5 inputs the missing 3D point cloud into the trained student model for inference and obtains the point cloud completion result.

[0157] To verify the effectiveness of this invention, we conducted comparative experiments on several typical point cloud completion models (including SeedFormer, AnchorFormer, and CRA-PCN) to evaluate its completion performance on three types of objects: cars, tables, and lamps. The evaluation metric was the average L1 Chamfer Distance × 10⁻⁶. 3 (Avg), the lower the value, the higher the completion accuracy. The table below shows the performance comparison of each model with and without knowledge distillation (KD) and different distillation methods:

[0158] Table 1. Performance comparison of different knowledge distillation methods on point cloud completion tasks.

[0159]

[0160] CasPoinTr innovatively breaks down the point cloud completion task into two collaborative stages: shape reconstruction and fusion completion. In the fusion completion stage, knowledge distillation is introduced, where a teacher model trained on a denser point cloud transfers "incomplete-complete" relational knowledge to the student model. This design helps the student model better grasp the overall shape of the object and accurately predict missing regions in the point cloud. In experiments on the ShapeNet-55 dataset, this method outperforms existing solutions in both shape recovery and detail preservation.

[0161] SCPNet is a point cloud-based semantic scene completion method that proposes three core solutions to improve model performance. One of these is a teacher-student knowledge distillation strategy, which transfers the dense semantic knowledge from multi-frame models to single-frame models. This approach ensures the efficiency of single-frame models while using distillation to compensate for the lack of semantic information in single-frame data. Furthermore, combined with a redesigned multi-path feature aggregation completion sub-network, the semantic scene completion effect is further enhanced.

[0162] The experimental results above demonstrate that the attention-guided anti-masking bi-branch distillation method proposed in this invention achieves optimal completion accuracy across multiple models and different object categories. Its performance consistently outperforms the student model without distillation (S w / oKD) and two point cloud completion methods (CasPoinTr and SCPNet). This fully validates that the proposed method, through attention-guided anti-masking, bi-branch collaborative learning, and triple-supervised optimization, can more effectively transfer knowledge from the teacher model, significantly improving the lightweight model's ability to restore geometric details and its overall robustness in complex scenes. This provides a reliable solution for deploying high-precision point cloud completion models in resource-constrained autonomous driving platforms.

[0163] It should be noted that the present invention is not limited to the above-described embodiments. The above embodiments are merely examples, and any embodiments that have the same structure and perform the same effects as the technical concept within the scope of the present invention are included within the scope of the present invention. Furthermore, various modifications that can be conceived by those skilled in the art to the embodiments, and other ways of constructing by combining some of the constituent elements of the embodiments, without departing from the spirit of the present invention, are also included within the scope of the present invention.

Claims

1. An attentional anti-masked dual-branch distillation method for three-dimensional point cloud completion, characterized in that, The method comprises the following steps: S1, extracting a multi-scale spatial attention map from a teacher model, and generating an anti-mask point cloud based on the attention map, specifically comprising: S11, extracting multi-scale feature maps from the shallow, middle and deep layers of the teacher model, and calculating the L2 norm of each feature map in the channel dimension to generate a multi-level spatial attention map; S12, upsampling the multi-level spatial attention map to a uniform size and weighting and fusing according to the importance weight of each level to obtain a fused attention map, and then performing normalization processing; S13, based on the normalized fused attention map, applying a dynamic anti-mask rule to generate a binary mask matrix, the dynamic anti-mask rule specifically being: For each position in the normalized fused attention map, if its normalized attention value is lower than a preset threshold, the position information is retained; If its normalized attention value is higher than or equal to the preset threshold, it is decided according to a random probability whether to retain or discard the position information; S14, copying the binary mask matrix along the channel dimension and element-wise multiplying it with the intermediate feature point cloud learned by the student model to generate an anti-mask point cloud; S2, constructing a double-branch student network, wherein the first branch processes the intermediate feature point cloud of the student model, and the second branch processes the anti-mask point cloud, and the specific structure of the double-branch student network is: The first branch is a standard distillation path, and the input thereof is the intermediate feature point cloud extracted from the student model backbone network; The second branch is an anti-mask distillation path, and the input thereof is the anti-mask point cloud, and the path includes a simple convolution generator for mapping the features of the anti-mask point cloud to the complete semantic space of the teacher model; S3, using a joint recursive distillation module to perform recursive upsampling and feature aggregation of the feature point clouds of the first branch, the second branch and the teacher model under geometric consistency constraint, the joint recursive distillation module adopts a recursive structure and performs feature upsampling and aggregation through a multi-scale point cloud set and a Channel-aware Transformer operation; S4, using a triple supervision optimization module to jointly optimize the CD distance loss, the feature distillation loss and the cross-branch contrast loss to train the student model; S5, inputting the missing three-dimensional point cloud into the trained student model for inference to obtain a point cloud completion result.

2. The attentional anti-masked dual-branch distillation method for 3D point cloud completion according to claim 1, characterized in that, The S3 specifically comprises: S31, performing hierarchical downsampling on the input point cloud by a farthest point sampling algorithm to obtain a point cloud set containing m scales; S32, at the m-1 scale, performing a Channel-aware Transformer operation to aggregate support information from the shape point cloud and the trajectory point cloud, and output enhanced point features; S33, concatenating the output features of the m-1 scale and the query features of the m-2 scale, performing fusion through a multi-layer perception, and generating a value vector for the m-2 scale CT operation; S34, continuing to perform the CT operation at the m-2 scale, and iteratively performing the process until high-level point cloud features are obtained.

3. The attentional anti-masked dual-branch distillation method for 3D point cloud completion according to claim 2, characterized in that, The specific process of the Channel-aware Transformer operation comprises: According to the query point, the key point and the value point, an attention weight is calculated through a vector attention mechanism; Based on the attention weight, a plurality of attention heads are used to aggregate local features to generate an up-sampling point feature; After the up-sampling point feature and the query feature are concatenated, local average pooling and weight calculation are performed for weighted fusion and addition of a residual connection to update the up-sampling point feature.

4. The attentional anti-masked dual-branch distillation method for 3D point cloud completion according to claim 1, characterized in that, The loss function used by the triple supervision optimization module in S4 includes: CD distance loss for directly optimizing the geometric error between the student model completion output and the real point cloud; Feature distillation loss for minimizing the difference between the output features of the first branch and the second branch of the student model and the output features of the teacher model; Cross-branch contrast loss for constraining the consistency of the first branch and the second branch of the student model in the output feature space.

Images