Point cloud self-supervised learning method based on multi-feature perception and auxiliary reconstruction

Abstract

The invention relates to the technical field of three-dimensional sensing, in particular to a point cloud self-supervised learning method based on multi-feature sensing and auxiliary reconstruction. The method comprises the following steps: S1, grouping point clouds of a random mask strategy; s2, local mark multi-feature fusion embedding is carried out; s3, pre-training the backbone network of the auto-encoder; s4, performing GaussPoint enhancement in a fine tuning stage; s5, auxiliary branch reconstruction in a fine tuning stage; and S6, carrying out multi-task coverage classification segmentation evaluation. According to the method, a Point-MAR network model is constructed on the basis of a mask self-encoding normal form by combining a multi-feature perception fusion embedder, an auxiliary reconstruction branch and a GaussPoint data enhancement method, feature expression, geometric perception and generalization capabilities are remarkably improved, and the method shows good competitiveness compared with numerous mainstream models in downstream tasks.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional sensing technology, specifically to a point cloud self-supervised learning method based on multi-feature perception and assisted reconstruction. Background Technology

[0002] With the rapid development of 3D sensing technology, point cloud data acquisition has become increasingly convenient. As an important data form representing 3D geometric structures, point clouds have been widely used in various fields such as autonomous driving, robot navigation, smart cities, and healthcare. Unlike 2D images, point cloud data is characterized by disorder, sparsity, and irregularity, posing significant challenges to deep neural network processing and effective point cloud feature extraction. Some works project point clouds into multi-view 2D images and utilize mature 2D convolutional networks to extract point cloud features, or discretize point cloud data into standardized 3D voxel grids and apply 3D convolutional neural networks (CNNs) for shape classification in volume representation. However, projection into 2D images or voxelization suffers from drawbacks such as potential loss of geometric information and high computational costs. PointNet, on the other hand, directly processes the original point cloud data, using max pooling to extract global features from all points. In the segmentation task, the global features are concatenated with the local features of each point, and then point-by-point prediction is performed using a multilayer perceptron (MLP), initiating extensive research on point cloud feature extraction methods. While the rapid development of 3D sensing technology has significantly reduced the cost and greatly improved the efficiency of point cloud data acquisition, high-quality, large-scale point cloud datasets remain relatively scarce compared to the massive and mature 2D image datasets. Furthermore, the disordered nature and sheer volume of point clouds make annotation difficult and expensive, severely hindering the widespread adoption of supervised learning methods for point cloud tasks. This has prompted researchers to turn to semi-supervised and self-supervised learning of point clouds, aiming to reduce reliance on labeled data while fully exploring the potential representations and semantic information of point clouds.

[0003] In recent years, self-supervised learning (SSL) methods for point clouds based on the Transformer mask autoencoder (MAE) paradigm have gradually become a research hotspot. For example, Point-MAE utilizes an asymmetric Transformer encoder-decoder architecture to effectively capture long-range dependencies in point clouds and learn robust global context representations. It employs a pre-training paradigm that divides the point cloud into irregular point blocks and performs mask reconstruction on high-proportion regions, learning latent feature representations from a large amount of unlabeled point cloud data, significantly improving the performance of downstream tasks. These methods fully demonstrate the potential and effectiveness of the standard Transformer in 3D point cloud modeling. Despite the significant success of Point-MAE, it still has the following three shortcomings. First, the geometric prior is not fully utilized in the feature extraction stage. It relies solely on local semantic feature extraction based on PointNet. While its standard Transformer architecture directly models patch tokens globally and captures long-distance dependencies through self-attention, it lacks effective modeling of the geometric structure of the local neighborhood of the point cloud. Furthermore, due to the inherent sparsity and irregularity of point cloud data, the feature distribution varies significantly between different local point blocks, making it impossible for a single feature channel to fully capture the rich geometric details and structural relationships within the point cloud, thus limiting the model's deep representation capabilities. Second, in the fine-tuning stage, there is a lack of explicit constraints between the pre-training task and the downstream task, ignoring the insufficient correlation between the downstream task and the pre-training task. Typically, only the cross-entropy loss of the downstream task is optimized, failing to effectively utilize the potential task relevance during pre-training. This easily leads to overfitting on limited labeled data and forgetting the general representations learned in pre-training. In particular, point clouds in different datasets and task environments vary significantly in noise levels, density distribution, and geometric complexity, thus limiting the model's generalization performance across datasets and multi-task scenarios. Summary of the Invention

[0004] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide a point cloud self-supervised learning method based on multi-feature perception and assisted reconstruction. By combining a multi-feature perception fusion embedder, an assisted reconstruction branch and the GaussPoint data augmentation method, a Point-MAR network model is constructed, which significantly improves feature representation, geometric perception and generalization ability.

[0005] The technical solution adopted in this invention is as follows:

[0006] A self-supervised learning method for point clouds based on multi-feature perception and assisted reconstruction includes the following steps:

[0007] S1. Point cloud grouping using random masking strategy: The point cloud is divided into local point blocks using FPS and KNN, and a high proportion of random masking is introduced to enhance the self-supervised learning ability.

[0008] S2, Local Label Multi-Feature Fusion Embedding: Define a multi-feature perceptual fusion embedding, extract local semantic, explicit geometric and affine features in parallel, and enhance feature representation through channel attention adaptive fusion;

[0009] S3. Autoencoder backbone network pre-training: Based on the standard Transformer block, an asymmetric encoder-decoder setup is adopted to process visible, encoded visible, and learnable mask tokens to carry out the pre-trained mask reconstruction task.

[0010] S4. GaussPoint Enhancement in the Fine-tuning Stage: A GaussPoint data augmentation method is proposed, which generates smooth geometric deformations based on periodic Gaussian functions to improve the robustness and generalization ability of the model.

[0011] S5, Fine-tuning stage assisted reconstruction branch: The fine-tuning stage introduces a geometry-aware generator, which uses relative position encoding and random initialization of its weights to reconstruct the point cloud to enhance geometric feature learning and alleviate the forgetting problem;

[0012] S6. Multi-task evaluation covering classification and segmentation: Object classification tasks cover classification of real object datasets / clean object datasets; the few-shot learning dataset includes four settings, reporting average accuracy and standard deviation; object partial segmentation uses the average intersection-union ratio of classes / instances as the evaluation metric.

[0013] This technical solution is based on the masked autoencoder paradigm. It extracts multiple features in parallel and adaptively fuses them through a multi-feature perceptual fusion embedder to enhance feature representation. The asymmetric encoder-decoder setup reduces computational overhead while promoting discriminative feature learning. The GaussPoint data augmentation method generates smooth geometric deformations, improving the model's robustness and generalization ability. In the fine-tuning stage, an auxiliary reconstruction branch is introduced, using relative position encoding and randomly initializing the weights of the geometry-aware generator to reconstruct the point cloud and enhance geometric feature learning.

[0014] In addition, the point cloud self-supervised learning method based on multi-feature perception and assisted reconstruction proposed according to the present invention also has the following additional technical features:

[0015] According to an embodiment of the present invention, the point cloud grouping of the random masking strategy in step S1 includes the following steps:

[0016] S11. Input point cloud is processed using FPS and KNN algorithms. Divided into a series of local points ;

[0017] For a containing Input point cloud of points The FPS algorithm is used to extract the input point cloud. Medium sampling Points are used as the center point set C. Based on the center point set C, the KNN algorithm is used to input the point cloud. Retrieve the k nearest neighbors of each center point to construct... Local point block P:

[0018] ,

[0019] ,

[0020] In the formula: The number of neighboring points contained in each local point block; the points within each local point block are represented as relative coordinates through a centering operation;

[0021] S12. Introduce a high-proportion random masking strategy at the local point block level;

[0022] Local point blocks are encoded into a sequence of tokens using an embedder. :

[0023] ,

[0024] Let the mask ratio be... , Classified as visible tokens With mask tokens :

[0025] ,

[0026] Visible tokens The input encoder extracts high-level contextual features, while the mask tokens... Use shared, learnable masktokens instead as input to the decoder.

[0027] According to an embodiment of the present invention, the local label multi-feature fusion embedding in step S2 includes the following steps:

[0028] The S21 and Point-MAE algorithms rely on a lightweight PointNet encoder to generate tokens representations of local point blocks. The encoder extracts local semantic features through shared MLP and pooling aggregation.

[0029] S22. The multi-feature perceptual fusion embedder is defined based on the PointNet encoder using the following features:

[0030] S221, Local Semantic Features: Given a block of points:

[0031] ,

[0032] Extracting local semantic features using a lightweight PointNet encoder Then, global semantic features are obtained through max pooling aggregation. :

[0033] ,

[0034] In the formula: For shared MLP convolutional layers; These are learnable parameters;

[0035] S222, Explicit Geometric Features: Through a geometric mapping network Explicitly encode the center point, neighboring points, and their relative positions to generate explicit geometric features. Injecting local geometric priors into the model:

[0036]

[0037] In the formula: The coordinates of the center point; The coordinates of the neighboring points; This refers to relative displacement; This is point-by-point multiplication;

[0038] S223, Affine Transformation Features: Introducing a Geometric Affine Module Local feature tensor Standardization and affine transformation are performed to alleviate the problem of inconsistent distribution among local points and enhance the robustness of features:

[0039]

[0040] In the formula: and These are the mean accuracy and standard deviation, respectively. As a stabilizing factor; For the affine parameters introduced;

[0041] S224, Channel Attention Fusion: Adaptively fuse three types of features through a channel attention module.

[0042]

[0043]

[0044] In the formula: It is a fully connected network; The learned fusion weights; Embedding of each point block output by the multi-feature-aware fusion embedder gather.

[0045] According to an embodiment of the present invention, the pre-training of the autoencoder backbone network in step S3 includes the following steps:

[0046] S31. The autoencoder backbone network is configured with an asymmetric encoder-decoder based on the standard Transformer block; the encoder only processes visible tokens. And add a visible center position embedding to each Transformer block. To provide location information, the encoded tokens are denoted as ;

[0047] S32, The decoder's Transformer block with visible tokens and mask tokens As input, and add complete position embeddings in each Transformer block. Provides location information for all tokens;

[0048] The encoder-decoder structure is described as follows:

[0049] ,

[0050] ,

[0051] In the formula: This refers to the mask ratio; This represents the number of local points / blocks. For the embedded dimension;

[0052] S33, Decoder Output Achieving reconstruction goals through prediction heads:

[0053] The prediction head uses a lightweight fully connected (FC) layer to project the features into a vector with the same dimensions as the total number of local point patch coordinates. Subsequently, the predicted masked local point patches are generated by reshaping. :

[0054] , .

[0055] According to one embodiment of the present invention, in the pre-training of the autoencoder backbone network in step S3, the model is pre-trained based on the ShapeNet dataset. ShapeNet covers 55 object categories and contains 51,300 3D models. The data is sampled at 1024 points using the farthest point sampling FPS and divided into 64 point blocks, each containing 32 points. These are embedded as local point block features and randomly masked with a mask ratio r=0.6 before being input into the Transformer for modeling. The encoder consists of 12 Transformer layers, and the decoder consists of 4 Transformer layers. Each Transformer block has a feature dimension of 384 and contains 6 self-attention heads. The optimizer uses AdamW with an initial learning rate of 0.001, a weight decay coefficient of 0.05, and a cosine annealing (CosLR) strategy for learning rate scheduling. The data augmentation method in the pre-training stage only uses random scaling and translation.

[0056] According to an embodiment of the present invention, the GaussPoint enhancement in the fine-tuning stage of step S4 includes the following steps:

[0057] S41 and GaussPoint utilize a periodic Gaussian function as a residual mapping to apply smooth and periodic geometric deformation to the input point cloud while maintaining topological consistency.

[0058] Define a periodic function as:

[0059]

[0060] In the formula: The width of the Gaussian kernel controls the smoothness of the deformation; The period of the function; To define the cutoff range, let... To balance computational efficiency with geometric continuity;

[0061] S42, Global Deformation GaussPoint-SGL: First, the point cloud is normalized to the unit sphere space, and then a single periodic Gaussian function is used to globally deform the entire point cloud; given the input point cloud... The enhanced point cloud is represented as follows:

[0062]

[0063] In the formula: The deformation amplitude; For frequency factors; These are phase shifts, all of which are randomly generated.

[0064] S43, Local Deformation GaussPoint-MUL: This mechanism uses a multi-anchor-point weighting method to generate geometric deformation in a localized area of ​​the point cloud, simulating subtle geometric variations on the object's surface; it selects M anchor points using either farthest-point sampling or random sampling. For each point, calculate its average offset to all anchor points. And based on this offset Constructing a local displacement field :

[0065]

[0066]

[0067] S44. Adaptive Smoothing Mechanism: GaussPoint adaptive smoothing mechanism, which uses the standard deviation of the input point cloud... Adjust the deformation range and Gaussian kernel width This ensures that the deformation process is smooth and proportional to the geometric scale.

[0068] .

[0069] According to an embodiment of the present invention, in the GaussPoint enhancement of the fine-tuning stage in step S4, the fine-tuning model adopts the multi-feature perceptual fusion embedder and encoder of the pre-training stage, adds a classification / segmentation head and introduces an auxiliary reconstruction branch after the encoder, and performs feature extraction and discrimination for classification or segmentation tasks.

[0070] According to an embodiment of the present invention, the fine-tuning stage of the reconstruction branch in step S5 includes the following steps:

[0071] S51. In the fine-tuning stage, a geometry-aware generator is introduced as an auxiliary reconstruction branch. The geometry-aware generator is based on the Transformer Decoder architecture, uses Relative Position Encoding (RPE) as input, and models relative orientation cues between groups to capture the geometric topological relationships between point cloud groups.

[0072] For the set of group centers:

[0073]

[0074] In the formula: For the first The coordinates of the center point of each group; The number of groups;

[0075] Relative Position Encoding (RPE) captures the relative geometric relationships of point clouds by calculating the normalized direction vector between adjacent group centers.

[0076] No. Groups Compared to the first Groups The unit direction vector is defined as:

[0077] ,

[0078]

[0079] In the formula: Use absolute center coordinates for the first group; To avoid small values ​​when divided by zero;

[0080] S52, The generator uses the output features of the encoder. With relative position encoding (RPE) as input, the generator's output is finally mapped to point cloud coordinates via a simple prediction head:

[0081] , .

[0082] According to an embodiment of the present invention, step S6, which involves multi-task evaluation of classification and segmentation, includes the following steps:

[0083] The optimization objective of fine-tuning the model is the loss of downstream tasks. With reconstruction losses The weighted combination, namely:

[0084]

[0085]

[0086] In the formula: and These are the predicted and true point sets, respectively. To balance the hyperparameters, the contribution of each loss is balanced.

[0087] According to one embodiment of the present invention, the multi-task aspect of step S6 covers the downstream tasks of evaluating model performance in classification and segmentation evaluation, including:

[0088] Object classification: divided into real-world datasets and clean object datasets; real-world datasets include OBJ_BG, OBJ_ONLY, and PB_T50_RS;

[0089] Few-shot learning: The dataset uses an n-way, m-shot configuration, where n is the number of classes randomly selected from the class set, and m is the number of samples randomly selected from each class. Experiments cover four configurations: {5-way, 10-shot}, {5-way, 20-shot}, {10-way, 10-shot}, and {10-way, 20-shot}. Each experiment is run independently for 10 runs, and the mean accuracy and standard deviation are reported.

[0090] Object Partial Segmentation: Employs a lightweight segmentation header consistent with Point-MAE, containing the average intersection-union ratio (mIoU) over all instance mIoU and all class mIoU.

[0091] Compared with the prior art, the present invention has the following advantages:

[0092] (1) Enhanced feature representation: Local semantic, explicit geometric and affine features are extracted in parallel by a multi-feature perception fusion embedder, and adaptive fusion is performed using a channel attention module to generate more robust and discriminative embedding tokens, which effectively improves the feature representation capability of the model.

[0093] (2) Enhance geometric perception: In the fine-tuning stage, a geometric perception generator is introduced, using relative position encoding and random initialization of its weights to reconstruct the point cloud and enhance geometric feature learning, effectively alleviating the forgetting problem;

[0094] (3) Enhance model robustness: The GaussPoint data augmentation method generates smooth geometric deformations through periodic Gaussian functions, maintains topological consistency, and improves the robustness and generalization ability of the model. Attached Figure Description

[0095] Figure 1 This is a flowchart illustrating the principle of the method of the present invention.

[0096] Figure 2 This is the overall architecture diagram of the Point-MAR of the present invention.

[0097] Figure 3 This is an architecture diagram of the multi-feature perception fusion embedder of the present invention.

[0098] Figure 4 Different amplitudes in GaussPoint-MUL mode Value and frequency factor The result of geometric deformation of the value.

[0099] Figure 5 This is a reconstruction result on the ShapeNet dataset during the pre-training phase.

[0100] Figure 6 The graph shows the results of parameter tuning experiments on the ScanObjectNN datasets OBJ_BG and OBJ_ONLY. Detailed Implementation

[0101] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0102] Example 1:

[0103] like Figure 1 As shown, this embodiment provides a self-supervised learning method for point clouds based on multi-feature perception and assisted reconstruction, including the following steps:

[0104] S1. Point cloud grouping using random masking strategy: The point cloud is divided into local point blocks using FPS and KNN, and a high proportion of random masking is introduced to enhance the self-supervised learning ability.

[0105] S2, Local Label Multi-Feature Fusion Embedding: Define a multi-feature perceptual fusion embedding, extract local semantic, explicit geometric and affine features in parallel, and enhance feature representation through channel attention adaptive fusion;

[0106] S3. Autoencoder backbone network pre-training: Based on the standard Transformer block, an asymmetric encoder-decoder setup is adopted to process visible, encoded visible, and learnable mask tokens to carry out the pre-trained mask reconstruction task.

[0107] S4. GaussPoint Enhancement in the Fine-tuning Stage: A GaussPoint data augmentation method is proposed, which generates smooth geometric deformations based on periodic Gaussian functions to improve the robustness and generalization ability of the model.

[0108] S5, Fine-tuning stage assisted reconstruction branch: The fine-tuning stage introduces a geometry-aware generator, which uses relative position encoding and random initialization of its weights to reconstruct the point cloud to enhance geometric feature learning and alleviate the forgetting problem;

[0109] S6. Multi-task evaluation covering classification and segmentation: Object classification tasks cover classification of real object datasets / clean object datasets; the few-shot learning dataset includes four settings, reporting average accuracy and standard deviation; object partial segmentation uses the average intersection-union ratio of classes / instances as the evaluation metric.

[0110] This technical solution is based on the masked autoencoder paradigm. It extracts multiple features in parallel and adaptively fuses them through a multi-feature perceptual fusion embedder to enhance feature representation. The asymmetric encoder-decoder setup reduces computational overhead while promoting discriminative feature learning. The GaussPoint data augmentation method generates smooth geometric deformations, improving the model's robustness and generalization ability. In the fine-tuning stage, an auxiliary reconstruction branch is introduced, using relative position encoding and randomly initializing the weights of the geometry-aware generator to reconstruct the point cloud and enhance geometric feature learning.

[0111] like Figure 2 As shown, firstly, the input point cloud is divided into a series of local point blocks using farthest point sampling (FPS) and K-nearest neighbor search (KNN). A multi-feature-aware fusion embedder extracts and integrates three types of complementary features from these local point blocks: local semantic features generated by convolution, explicit geometric features, and normalized affine features. An adaptive weighted fusion of multi-branch features is achieved through a channel attention fusion module, generating more robust and discriminative embedding tokens. These tokens are input to a Transformer encoder to capture global dependencies, and then the decoder performs point cloud pre-training reconstruction. In the fine-tuning stage, a geometry-aware generator is introduced as an auxiliary reconstruction branch to reconstruct the complete point cloud. This is jointly optimized with downstream tasks, i.e., cross-entropy loss combined with Chamfer distance, to enhance the geometric representation of features and alleviate forgetting and other problems. Furthermore, this invention designs a GaussPoint data augmentation method in the fine-tuning stage, which performs smooth geometric transformations on the point cloud data based on a periodic Gaussian function while maintaining topological consistency, further improving the model's robustness and generalization ability.

[0112] To address the challenges of large scale, unordered nature, and sparsity in point cloud data, the input point cloud is divided into a series of local point patches using Farthest Point Sampling (FPS) and K-Nearest Neighbors (KNN). Specifically: for a point cloud containing... Input point cloud of points First, through FPS from Medium sampling Points are used as the center point set C. Then, based on the center point set, the KNN algorithm is used to retrieve the k nearest neighbors of each center point in the input point cloud X, thereby constructing... Local point block P:

[0113] ,

[0114] ,

[0115] In the formula: The number of neighborhood points contained in each local point block. Points within each local point block are further represented as relative coordinates through a centering operation, which significantly enhances the model's ability to model local geometry and reduces the impact of location information leakage.

[0116] To enhance the model's self-supervised learning capability, a high-proportion random masking strategy is introduced at the local point block level. Local point blocks are encoded into a sequence of tokens using an embedder. , Specifically, let the mask ratio be... , Classified as visible tokens With mask tokens :

[0117] ,

[0118] In subsequent stages, visible tokens are input into the Transformer encoder to extract high-level contextual features, while the masking... Use shared, learnable mask tokens instead as input to the decoder.

[0119] Point-MAE relies on a lightweight PointNet-based encoder to generate token representations of local point blocks. This encoder extracts local semantic features through shared MLP and pooling operations. This mechanism effectively captures high-level semantic information of point clouds, providing a foundation for Transformer to model long-distance dependencies and global self-attention in subsequent stages. However, a single semantic channel is still insufficient for modeling the explicit local geometry of point clouds, while geometric information is the most fundamental prior knowledge in point cloud data. Furthermore, due to the sparsity and irregularity of local regions in point clouds, significant geometric differences exist between different regions. Relying solely on a single semantic feature extractor often fails to maintain the stability and generalization of representations across all scenarios. Figure 3 As shown, the multi-feature perception fusion embedder architecture extracts deep semantic features, explicit geometric features, and standardized affine features in parallel from local point blocks. Then, it uses the channel attention fusion module to achieve adaptive feature fusion, where B is the batch size and C is the feature dimension.

[0120] The multi-feature-aware fusion embedder, based on the PointNet encoder, introduces explicit geometric features and normalized affine features, and designs a channel attention fusion module to adaptively weight and fuse the three-branch features, thereby generating more robust local point-block embedding tokens with stronger geometric feature representation capabilities. Specifically:

[0121] Local semantic features: given point blocks , First, local semantic features are extracted using a lightweight PointNet, and then global semantic features are obtained by max pooling.

[0122] ,

[0123] In the formula: It is a shared MLP convolutional layer. These are learnable parameters.

[0124] Explicit geometric features: To capture the explicit geometric relationships of local point blocks, a geometric mapping network is designed. Explicitly encoding the center point, neighboring points, and their relative positions generates explicit geometric features, injecting essential prior knowledge of local geometric structure into the model:

[0125]

[0126] In the formula: The coordinates of the center point The coordinates of the neighboring points and their relative displacements , This indicates point-by-point multiplication.

[0127] Affine Transformation Features: Inspired by PointMLP feature normalization, a geometric affine module is introduced to transform local feature tensors. Standardization and affine transformation are performed to effectively alleviate the problem of inconsistent distribution among local points and enhance the robustness of features.

[0128] , ,

[0129] In the formula: and These are the mean accuracy and standard deviation, respectively. It is a stabilizing factor. The affine parameters are introduced.

[0130] Channel attention fusion: To fully utilize the three types of features, a channel attention module is provided to adaptively fuse the three types of features.

[0131]

[0132]

[0133] In the formula: For a fully connected network, For the learned fusion weights, Embedding of each point block output by the multi-feature-aware fusion embedder gather.

[0134] Drawing inspiration from Point-MAE, the autoencoder backbone network is entirely based on standard Transformer blocks, employing an asymmetric encoder-decoder design. The encoder only processes visible tokens. And add a visible center position embedding to each Transformer block. To provide location information, the encoded tokens are denoted as The decoder has a smaller number of Transformer blocks, making visible tokens more common. and mask tokens As input, and add complete position embeddings in each Transformer block. This provides location information for all tokens. The encoder-decoder structure can be described as follows:

[0135] ,

[0136] ,

[0137] In the formula: For the mask ratio, This represents the number of local points and blocks. For the embedding dimension. By moving the mask tokens to the lightweight decoder, this setup reduces the encoder's computational overhead and avoids premature leakage of positional information, promoting the encoder's learning of more discriminative latent features. Finally, the decoder's output... The reconstruction goal is achieved through a simple prediction head: the prediction head uses a lightweight fully connected FC layer to project the features into a vector with the same dimensions as the total number of local point patch coordinates, and then reshapes the features to generate the predicted mask of local point patches.

[0138] , .

[0139] In the fine-tuning phase, the model retains the multi-feature perceptual fusion embedder and encoder from the pre-training phase. A classification / segmentation head and an auxiliary reconstruction branch are added after the encoder, allowing the model to extract and discriminate features for classification or segmentation tasks. However, the supervision signal for downstream tasks relies solely on class labels and lacks constraints on the geometric details of the point cloud, which can easily lead to overfitting or forgetting the general geometric representations learned in the pre-training phase.

[0140] To address these issues and improve the model's effective utilization of general representations learned during pre-training, a lightweight geometry-aware generator is introduced as an auxiliary reconstruction branch during the fine-tuning stage, such as... Figure 2 As shown in the diagram, this generator is based on the Transformer Decoder architecture. Unlike the pre-training stage, the generator uses Relative Position Encoding (RPE) as input instead of directly exposing the absolute center coordinates of local point blocks in the mask. It also models relative orientation cues between groups to accurately capture the geometric topological relationships between point cloud groups, such as relative orientation and adjacency. Simultaneously, the generator does not use the weights of the pre-trained decoder. The relative position encoding and random initialization of the generator's weights significantly increase the challenge of the reconstruction task, preventing the reconstruction task from taking "shortcuts." The generator may complete local coordinate reconstruction by directly utilizing absolute position information and pre-trained weights. This forces the encoder to retain and enhance its representation of local and global geometric structures, thereby learning more general and robust geometric features. Through a dual-task collaborative optimization mechanism of "classification (segmentation) - reconstruction," the model benefits from additional geometric reconstruction signals while receiving class supervision, effectively suppressing overfitting, reducing representation forgetting, and prompting the model to retain the general geometric features learned during pre-training.

[0141] Specifically, for the set of group centers ( For the first The coordinates of the center point of each group To determine the number of groupings, RPE captures the relative geometric relationships of the point cloud by calculating the normalized direction vectors between the centers of adjacent groups; the... The group relative to the first The unit direction vector of each group is defined as:

[0142] ,

[0143]

[0144] In the formula: Use absolute center coordinates for the first group. To avoid small values ​​that are divided by zero, the generator uses the encoder's output features. With relative position encoding as input, the generator's output is finally mapped to point cloud coordinates through a simple prediction head:

[0145] ,

[0146] Data augmentation is a key method to address data scarcity and improve model robustness and generalization ability. However, due to the disorder, sparsity and irregularity of point clouds, traditional augmentation strategies (such as adding random noise or rigid transformations) often fail to enrich geometric diversity and are prone to disrupting the topological consistency of point clouds, resulting in unnatural distortions or semantic biases in the generated samples (such as abrupt changes on the object surface or discontinuous structures), which limits the model's understanding of real-world scenes.

[0147] To address this, GaussPoint, a data augmentation method based on periodic Gaussian functions, is proposed. It effectively simulates the natural perturbation characteristics of real objects under surface undulations and curvature by generating smooth and physically consistent geometric deformations in the point cloud space. Furthermore, homeomorphic transformations ensure the topological consistency of the point cloud. Given the significant differences in noise levels, density distributions, and geometric complexity of point clouds across different datasets and task environments, GaussPoint includes two flexible variants to adapt to diverse application scenarios:

[0148] (1) Global Deformation GaussPoint-SGL is used to apply global, uniform geometric perturbation to the entire point cloud, thereby achieving consistent geometric deformation. It is suitable for synthetic datasets with low noise and regular shapes, such as ModelNet40, which rely more on global geometric consistency for shape recognition and classification.

[0149] (2) Local deformation GaussPoint-MUL is used to perform local deformation around multiple selected anchor points. The local Gaussian deformation guided by multiple anchor points enhances the model's local feature perception and noise adaptation capabilities. It is suitable for real-world scanning data with strong noise and occlusion features, such as ScanObjectNN, as well as part segmentation tasks with high requirements for local geometry, such as component boundaries and connections.

[0150] To more intuitively demonstrate the impact of GaussPoint on point cloud geometry, in Figure 4 The image shows ring point cloud data with different amplitudes in GaussPoint-MUL mode. Value and frequency factor The geometric deformation results of the value show that: when and When different values ​​are taken, the point cloud surface exhibits a continuous change from slight undulations to obvious fluctuations. and When the value is too large, it introduces severe, non-physical distortions, leading to deformation of the point cloud geometry. Figure 4 Displayed in a fixed Below, different The resulting deformation; shown below in the fixed Below, different The resulting deformation. When the amplitude or If the value is too large, the deformation will increase significantly.

[0151] The core of GaussPoint is to use a periodic Gaussian function as a residual mapping to apply smooth and periodic geometric deformation to the input point cloud while maintaining topological consistency. Specifically, the periodic function is defined as:

[0152]

[0153] In the formula: The width of the Gaussian kernel controls the smoothness of the deformation; For the period of the function, To define the cutoff range, it is usually set as follows: To balance computational efficiency with geometric continuity.

[0154] (1) Global Deformation GaussPoint-SGL: First, the point cloud is normalized to the unit sphere space, and then a single periodic Gaussian function is used to globally deform the entire point cloud. Given an input point cloud... The enhanced point cloud is represented as follows:

[0155]

[0156] In the formula: Indicates the deformation range. For frequency factors, The phase shifts are all randomly generated. This randomness design ensures that the deformation patterns vary independently for each batch and each dimension, generating a variety of enhanced samples that cover a wider range of geometric variations and prevent the model from overfitting to specific deformation patterns.

[0157] (2) Local Deformation GaussPoint-MUL: This method uses a multi-anchor-point weighting mechanism to generate geometric deformations in the local point cloud, simulating subtle geometric variations on the object's surface and enhancing the model's ability to represent the object's local geometric features. M anchor points are selected using either farthest-point sampling or random sampling. For each point, calculate its average offset to all anchor points. And based on this offset, a local displacement field is constructed:

[0158]

[0159]

[0160] (3) Adaptive Smoothing Mechanism: To adapt to the diversity of point cloud scales in real-world scenes, such as large objects having a wider deformation range and small objects having a narrower deformation range, GaussPoint sets up an adaptive smoothing mechanism, which uses the standard deviation of the input point cloud to achieve this smoothing. Adjust the deformation range and Gaussian kernel width This ensures that the deformation process is smooth and proportional to the geometric scale.

[0161]

[0162] The adaptive smoothing mechanism effectively avoids the unreasonable situation of "over-deformation of small objects and under-deformation of large objects," generating smoother and physically consistent augmented samples. To ensure fair comparison, this invention only uses the augmented point cloud for evaluation. Experimental results show that even with a single-fold increase in data volume, GaussPoint still effectively improves the robustness and generalization performance of the model, verifying its effectiveness in point cloud augmentation tasks.

[0163] During the pre-training phase, this invention uses L2 chamfer distance as the reconstruction loss to recover the coordinate information of the masked local point blocks. During the fine-tuning phase, the optimization objective of the model is the downstream task loss. With reconstruction losses The weighted combination, namely:

[0164]

[0165]

[0166] In the formula: and These represent the predicted and true point sets, respectively. To balance the hyperparameters, the contribution of each loss is balanced.

[0167] Example 2:

[0168] To verify the effectiveness of the method proposed in Example 1, the following experiments were conducted: (1) self-supervised pre-training on the ShapeNet dataset; (2) evaluating the model performance in various downstream tasks, including object classification, few-shot learning, and object partial segmentation; (3) ablation studies; and (4) GaussPoint generalization experiments.

[0169] The model was pre-trained on the ShapeNet dataset, which covers 55 common object categories and contains approximately 51,300 3D models. First, 1024 points were sampled using farthest point sampling (FPS) and divided into 64 point blocks (each block containing 32 points). These blocks were then embedded as local point block features and randomly masked with a mask ratio r=0.6 before being input into a Transformer for modeling. The encoder consisted of 12 Transformer layers, and the decoder consisted of 4 Transformer layers. Each Transformer block had a feature dimension of 384 and contained 6 self-attention heads. The optimizer used AdamW with an initial learning rate of 0.001, a weight decay coefficient of 0.05, and a cosine annealing (CosLR) strategy for learning rate scheduling. Following previous work, data augmentation during the pre-training phase only used random scaling and translation. An example of point cloud reconstruction during the pre-training phase is shown below. Figure 5 As shown, Point-MAR exhibits good reconstruction results, proving its effectiveness.

[0170] Real-world dataset object classification: The ScanObjectNN point cloud dataset from real indoor scene scans was used, covering 15 categories and a total of 15,000 objects. Three variants were provided: OBJ_BG (including background and objects), OBJ_ONLY (objects only), and PB_T50_RS (including point cloud perturbation, background, and random sampling). Experiments were conducted on these three variants. During training, data augmentation included rotation and the proposed GaussPoint-MUL method. The results are shown in Table 1, fully validating the effectiveness of the proposed Point-MAR model. Specifically, Point-MAR achieved accuracy improvements of 4.13%, 5.00%, and 4.06% compared to the baseline method Point-MAE on the OBJ_BG, OBJ_ONLY, and PB_T50_RS variants, respectively. Furthermore, compared to various improved methods based on the MAE framework and current leading multimodal self-supervised learning methods, Point-MAR still demonstrates significant performance advantages and strong competitiveness, proving its classification generalization ability and application potential in real-world complex scenes.

[0171] Table 1: Classification Accuracy of ScanObjectNN

[0172]

[0173] In Table 1, #P(M) reports the parameters of the inference model; * indicates the results reported in the original paper that use the same data augmentation as the pre-training phase of this invention; the best results are highlighted in bold.

[0174] Clean Object Dataset Object Classification: Experiments were conducted on the synthetic dataset ModelNet40, which is based on clean 3D CAD models and contains 12,311 3D models covering 40 common object categories. Data augmentation during training employed randomized scaling and translation, along with the GaussPoint-SGL method. To ensure fair comparison, a standard voting strategy was used during full-parameter fine-tuning testing, and the results are shown in Table 2. Furthermore, to demonstrate the effectiveness of Point-MAR, experiments were conducted under the MLP-Linear and MLP-3 protocols (i.e., freezing the pre-trained backbone and training only the lightweight classifier). Point-MAR still exhibits superior performance compared to the baseline method, indicating that the learned features possess good transferability and generality.

[0175] Table 2: Classification Accuracy Table for ModelNet Dataset

[0176]

[0177] In Table 2, [ST] represents the standard Transformer architecture; * represents the reproduction results using the official pre-trained weights.

[0178] Few-shot learning: Few-shot learning experiments were conducted on ModelNet40 using an n-way, m-shot configuration. In the formula, n represents the number of classes randomly selected from the class set, and m represents the number of samples randomly selected from each class. Experiments covered four common few-shot configurations: {5-way, 10-shot}, {5-way, 20-shot}, {10-way, 10-shot}, and {10-way, 20-shot}. Following standard protocols, 10 experiments were run independently for each configuration, and the mean accuracy and standard deviation were reported. The results are shown in Table 3. Point-MAR's performance is comparable to many mainstream algorithms, especially with a 1.1% improvement on the 5-way, 10-shot configuration. This fully demonstrates that the proposed model can quickly adapt to new class tasks relying only on a small number of labeled samples, validating its good transfer and generalization capabilities.

[0179] Table 3: Overall Accuracy of Few-Shot Classification Reports on ModelNet40

[0180]

[0181] Partial Object Segmentation: To verify the representation learning ability of Point-MAR in downstream tasks, partial segmentation experiments were conducted on the ShapeNetPart dataset. This dataset contains 16,881 3D shapes, covering 16 common object categories (such as chairs, tables, airplanes, etc.). Unlike classification tasks, part segmentation requires the model to not only identify the overall category of the object, but also accurately capture the local geometric details and semantic differences of the object, thus better reflecting the refinement ability of the model's learned representation. A lightweight segmentation head design consistent with Point-MAE was adopted, and the data augmentation strategy employed random scaling translation and the GaussPoint-MUL pattern to improve the model's robustness to local deformation and geometric awareness. The experimental results are shown in Table 4, reporting the mean intersection-union ratio (mIoU) across all instances (mIoU) and all classes (mIoU). The results show that Point-MAR achieves significantly better performance than the baseline methods, reaching 84.8% in Cls.mIoU, and improving upon Point-MAE by 0.6%. This performance significantly outperforms many mainstream single-modal and multimodal methods, demonstrating Point-MAR's outstanding advantages in capturing local geometry and enhancing fine-grained point cloud representation.

[0182] Table 4: Partial Segmentation Results on ShapeNetPart

[0183]

[0184] During the fine-tuning phase, the optimization objective of the model is defined as a weighted combination of the downstream task loss and the geometric reconstruction loss: This aims to enhance the model's ability to perceive point cloud structure while maintaining task discriminativeness. This is the trade-off factor between the two. If the value is too small, the model tends to optimize downstream tasks, which may result in insufficient geometric representation capabilities; while When the weight is too large, it may distract the model from discriminative features, reducing the model's ability to distinguish geometric distributions. Systematic experiments were conducted on two variants of the ScanObjectNN dataset, OBJ_BG and OBJ_ONLY, with different weight ratios. The experimental results are as follows: Figure 6 As shown, when When the value is 3, a balance can be struck between the two, enabling the model to maintain sensitivity to point cloud structure while achieving excellent performance in downstream tasks.

[0185] To verify the effectiveness of the Point-MAR method of this invention, a systematic ablation study was conducted on the ScanObjectNN dataset for the main components of this invention: the multi-feature perceptual fusion embedder, the fine-tuning stage-assisted reconstruction branch, and the GaussPoint data augmentation module. Specifically, the experimental results are shown in Table 5. The modules and methods proposed in this invention all significantly improve the model performance, fully verifying the effectiveness of each module.

[0186] Table 5: Ablation experiments of key components of Point-MAR on the ScanObjectNN dataset

[0187]

[0188] To prevent the generator from taking shortcuts when reconstructing point clouds, relative position encoding is used instead of absolute position as the generator input in the auxiliary reconstruction branch, and the generator weights are randomly initialized. This significantly increases the difficulty of reconstructing point clouds, forcing the encoder to accurately capture the geometric topological relationships between point cloud groups and learn more general geometric features. The experimental results in Table 6 show that performance degrades when using absolute position embeddings and pre-trained decoder weights.

[0189] Table 6: Impact of Relative Position Embedding and Random Initialization Generator Weights

[0190]

[0191] To fully verify the effectiveness and universality of the proposed GaussPoint data augmentation strategy, systematic experimental evaluations were conducted on several representative point cloud learning frameworks, with the data augmentation strategy consistent with that described above. Specifically, the experiments covered two main categories of mainstream models: one is fully supervised learning models, such as PointNet and PointNet++; the other is self-supervised pre-trained models, such as Point-GPT. The experimental results are shown in Table 7. Under different network structures and training paradigms, GaussPoint delivers stable and significant performance improvements, particularly on real-world datasets with backgrounds and objects. The experimental results demonstrate the effectiveness and general adaptability of the GaussPoint data augmentation method in improving the quality of point cloud representations and its generalization to downstream tasks.

[0192] Table 7: Generalization Experiment Table of GaussPoint Data Augmentation Method

[0193]

[0194] In summary, this invention proposes a self-supervised learning network model for point clouds—Point-MAR—based on the mask autoencoder paradigm. By combining a multi-feature-aware fusion embedder, an auxiliary reconstruction branch, and GaussPoint data augmentation, it significantly improves feature representation, geometric perception, and generalization capabilities. The multi-feature-aware fusion embedder effectively integrates local semantics, explicit geometry, and normalized affine features, enhancing embedding robustness. The randomly initialized auxiliary reconstruction branch mitigates forgetting and catastrophic forgetting problems through relative position embedding and joint optimization. GaussPoint data augmentation generates topologically consistent geometric deformations, improving the model's robustness and generalization ability. Experiments demonstrate that Point-MAR outperforms Point-MAE in downstream tasks such as classification and segmentation, and exhibits strong competitiveness compared to various mainstream advanced methods.

[0195] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the invention should also be covered within the protection scope of the invention. Therefore, the protection scope of the invention should be determined by the scope of the claims.

Claims

1. A self-supervised learning method for point clouds based on multi-feature perception and assisted reconstruction, characterized in that, Includes the following steps: S1. Point cloud grouping using random masking strategy: The point cloud is divided into local point blocks using FPS and KNN, and a high proportion of random masking is introduced to enhance the self-supervised learning ability. S2, Local Label Multi-Feature Fusion Embedding: Define a multi-feature perceptual fusion embedding, extract local semantic, explicit geometric and affine features in parallel, and enhance feature representation through channel attention adaptive fusion; S3. Autoencoder backbone network pre-training: Based on the standard Transformer block, an asymmetric encoder-decoder setup is adopted to process visible, encoded visible, and learnable mask tokens to carry out the pre-trained mask reconstruction task. S4. GaussPoint Enhancement in the Fine-tuning Stage: A GaussPoint data augmentation method is proposed, which generates smooth geometric deformations based on periodic Gaussian functions to improve the robustness and generalization ability of the model. S5, Fine-tuning stage assisted reconstruction branch: The fine-tuning stage introduces a geometry-aware generator, which uses relative position encoding and random initialization of its weights to reconstruct the point cloud to enhance geometric feature learning and alleviate the forgetting problem; S6. Multi-task coverage of classification and segmentation evaluation: Object classification tasks cover classification of real object datasets / clean object datasets; The few-shot learning dataset includes four settings, reporting average accuracy and standard deviation; object partial segmentation uses the average intersection-union ratio of categories / instances as the evaluation metric.

2. The self-supervised learning method for point clouds based on multi-feature perception and assisted reconstruction as described in claim 1, characterized in that, The point cloud grouping of the random masking strategy in step S1 includes the following steps: S11. Input point cloud is processed using FPS and KNN algorithms. Divided into a series of local points ; For a containing Input point cloud of points The FPS algorithm is used to extract the input point cloud. Medium sampling Points are used as the center point set C. Based on the center point set C, the KNN algorithm is used to input the point cloud. Retrieve the k nearest neighbors of each center point to construct... Local point block P: , , In the formula: The number of neighboring points contained in each local point block; the points within each local point block are represented as relative coordinates through a centering operation; S12. Introduce a high-proportion random masking strategy at the local point block level; Local point blocks are encoded into a sequence of tokens using an embedder. : ， Let the mask ratio be... , Classified as visible tokens With mask tokens : ， Visible tokens The input encoder extracts high-level contextual features, while the mask tokens... Use shared, learnable masktokens instead as input to the decoder.

3. The self-supervised learning method for point clouds based on multi-feature perception and assisted reconstruction as described in claim 2, characterized in that, The local label multi-feature fusion embedding in step S2 includes the following steps: S21 and Point-MAE algorithms rely on a lightweight PointNet encoder to generate tokens representations of local point blocks. The encoder extracts local semantic features through shared MLP and pooling aggregation. S22. The multi-feature perceptual fusion embedder is defined based on the PointNet encoder using the following features: S221, Local Semantic Features: Given a block of points: ， Extracting local semantic features using a lightweight PointNet encoder Then, global semantic features are obtained through max pooling aggregation. : ， In the formula: For shared MLP convolutional layers; These are learnable parameters; S222, Explicit Geometric Features: Through a geometric mapping network Explicitly encode the center point, neighboring points, and their relative positions to generate explicit geometric features. Injecting local geometric priors into the model: In the formula: The coordinates of the center point; The coordinates of the neighboring points; This refers to relative displacement; This is point-by-point multiplication; S223, Affine Transformation Features: Introducing a Geometric Affine Module Local feature tensor Standardization and affine transformation are performed to alleviate the problem of inconsistent distribution among local points and enhance the robustness of features: In the formula: and These are the mean accuracy and standard deviation, respectively. As a stabilizing factor; For the introduced affine parameters; S224, Channel Attention Fusion: Adaptively fuse three types of features through a channel attention module. In the formula: It is a fully connected network; The learned fusion weights; Embedding of each point block output by the multi-feature-aware fusion embedder gather.

4. The self-supervised learning method for point clouds based on multi-feature perception and assisted reconstruction as described in claim 3, characterized in that, The pre-training of the autoencoder backbone network in step S3 includes the following steps: S31. The autoencoder backbone network is configured with an asymmetric encoder-decoder based on the standard Transformer block; the encoder only processes visible tokens. And add a visible center position embedding to each Transformer block. To provide location information, the encoded tokens are denoted as ; S32, The decoder's Transformer block with visible tokens and mask tokens As input, and add complete position embeddings in each Transformer block. Provides location information for all tokens; The encoder-decoder structure is described as follows: ， ， In the formula: This refers to the mask ratio; This represents the number of local points / blocks. For the embedded dimension; S33, Decoder Output Achieving reconstruction goals through prediction heads: The prediction head uses a lightweight fully connected (FC) layer to project the features into a vector with the same dimensions as the total number of local point patch coordinates. Subsequently, the predicted masked local point patches are generated by reshaping. : ， 。 5. The self-supervised learning method for point clouds based on multi-feature perception and assisted reconstruction as described in claim 4, characterized in that, In the pre-training of the autoencoder backbone network in step S3, the model is pre-trained based on the ShapeNet dataset. ShapeNet covers 55 object categories and contains 51,300 3D models. The data is sampled at 1024 points using the farthest point sampling FPS and divided into 64 point blocks, each containing 32 points. These are embedded as local point block features and randomly masked with a mask ratio r=0.6 before being input into the Transformer for modeling. The encoder consists of 12 Transformer layers, and the decoder consists of 4 Transformer layers. Each Transformer block has a feature dimension of 384 and contains 6 self-attention heads. The optimizer uses AdamW with an initial learning rate of 0.001, a weight decay coefficient of 0.05, and a cosine annealing (CosLR) strategy for learning rate scheduling. The data augmentation method in the pre-training stage is only random scaling and translation.

6. The self-supervised learning method for point clouds based on multi-feature perception and assisted reconstruction as described in claim 4, characterized in that, The fine-tuning stage of GaussPoint enhancement in step S4 includes the following steps: S41 and GaussPoint utilize a periodic Gaussian function as a residual mapping to apply smooth and periodic geometric deformation to the input point cloud while maintaining topological consistency. Define a periodic function as: In the formula: The width of the Gaussian kernel controls the smoothness of the deformation; The period of the function; To define the cutoff range, let... To balance computational efficiency with geometric continuity; S42, Global Deformation GaussPoint-SGL: First, the point cloud is normalized to the unit sphere space, and then a single periodic Gaussian function is used to globally deform the entire point cloud; given the input point cloud... The enhanced point cloud is represented as follows: In the formula: The deformation amplitude; For frequency factors; These are phase shifts, all of which are randomly generated. S43, Local Deformation GaussPoint-MUL: This mechanism uses a multi-anchor-point weighting method to generate geometric deformation in a localized area of ​​the point cloud, simulating subtle geometric variations on the object's surface; it selects M anchor points using either farthest-point sampling or random sampling. For each point, calculate its average offset to all anchor points. And based on this offset Constructing a local displacement field : S44. Adaptive Smoothing Mechanism: GaussPoint adaptive smoothing mechanism, which uses the standard deviation of the input point cloud... Adjust the deformation range and Gaussian kernel width This ensures that the deformation process is smooth and proportional to the geometric scale. 。 7. The self-supervised learning method for point clouds based on multi-feature perception and assisted reconstruction as described in claim 6, characterized in that, In the GaussPoint enhancement stage of step S4, the fine-tuning model uses the multi-feature perceptual fusion embedder and encoder from the pre-training stage, adds a classification / segmentation head and introduces an auxiliary reconstruction branch after the encoder, and performs feature extraction and discrimination for classification or segmentation tasks.

8. The self-supervised learning method for point clouds based on multi-feature perception and assisted reconstruction as described in claim 6, characterized in that, The fine-tuning stage of the reconstruction branch in step S5 includes the following steps: S51. In the fine-tuning stage, a geometry-aware generator is introduced as an auxiliary reconstruction branch. The geometry-aware generator is based on the Transformer Decoder architecture, uses Relative Position Encoding (RPE) as input, and models relative orientation cues between groups to capture the geometric topological relationships between point cloud groups. For the set of group centers: In the formula: For the first The coordinates of the center point of each group; The number of groups; Relative Position Encoding (RPE) captures the relative geometric relationships of point clouds by calculating the normalized direction vector between adjacent group centers. No. Groups Compared to the first Groups The unit direction vector is defined as: ， In the formula: Use absolute center coordinates for the first group; To avoid small values ​​when divided by zero; S52, The generator uses the output features of the encoder. With relative position encoding (RPE) as input, the generator's output is finally mapped to point cloud coordinates via a simple prediction head: ， 。 9. The self-supervised learning method for point clouds based on multi-feature perception and assisted reconstruction as described in claim 8, characterized in that, The multi-task evaluation in step S6, which covers classification and segmentation, includes the following steps: The optimization objective of fine-tuning the model is the loss of downstream tasks. With reconstruction losses The weighted combination, namely: In the formula: and These are the predicted and true point sets, respectively. To balance the hyperparameters, the contribution of each loss is balanced.

10. The self-supervised learning method for point clouds based on multi-feature perception and assisted reconstruction as described in claim 9, characterized in that, The multi-task aspect of step S6, which covers classification and segmentation evaluation, includes downstream tasks that evaluate the model's performance, such as: Object classification: divided into real-world datasets and clean object datasets; real-world datasets include OBJ_BG, OBJ_ONLY, and PB_T50_RS; Few-shot learning: The dataset uses an n-way, m-shot configuration, where n is the number of classes randomly selected from the class set, and m is the number of samples randomly selected from each class. Experiments cover four configurations: {5-way, 10-shot}, {5-way, 20-shot}, {10-way, 10-shot}, and {10-way, 20-shot}. Each experiment is run independently for 10 runs, and the mean accuracy and standard deviation are reported. Object Partial Segmentation: Employs a lightweight segmentation header consistent with Point-MAE, containing the average intersection-union ratio (mIoU) over all instance mIoU and all class mIoU.

Images