Context point cloud modeling implementation method and device for weakly supervised point cloud segmentation

Abstract

The application discloses a context point cloud modeling implementation method and device for weakly supervised point cloud segmentation, relates to the technical field of context information understanding in cloud segmentation, and the method performs continuity masking on regional point cloud data, and constructs a mask feature prediction task which is beneficial to model learning point cloud data context; by constraining the consistency of corresponding features before / after the same point cloud data mask, the model can effectively utilize the neighborhood information in the mask data, improve the context understanding of the model for the point cloud data, and finally realize the performance of semantic segmentation of the point cloud segmentation model in the weak labeling scene. The application effectively solves the point cloud semantic segmentation problem in the scene where the labeling is very sparse and the point cloud structure is very complex, and can be widely used in large-scale 3D visual understanding application scenes.

Description

Technical Field

[0001] This invention relates to the technical field of contextual information understanding in cloud segmentation, and in particular to a method and apparatus for implementing contextual point cloud modeling for weakly supervised point cloud segmentation. Background Technology

[0002] With the increasing demand for autonomous driving and robotic navigation, point cloud semantic segmentation has become an indispensable technology for accurate 3D environment perception. In recent years, fully supervised learning for point cloud segmentation has made significant progress. However, the number of points in point cloud data can easily reach tens of thousands, making dense labeling of points both time-consuming, labor-intensive, and inefficient. Undoubtedly, in-depth research into point cloud semantic segmentation from sparse labels is crucial for reducing labeling costs and expanding application scenarios.

[0003] In recent years, to reduce reliance on dense labels while still delivering satisfactory point cloud semantic segmentation performance, most work has focused on learning from weakly labeled data. Among several types of weakly labeled data, partial labeling schemes offer the best balance between labeling cost and segmentation performance. In partially labeled point cloud data, the labeled portions typically represent only a small fraction (e.g., 0.1%) of the points in each scene. In this case, directly applying supervised cross-entropy loss function training to only the limited labeled portions can easily lead to model overfitting. Therefore, the main challenge of this task is to learn from a large number of unlabeled points to improve the model's generalization performance.

[0004] Existing methods address this challenge by leveraging feature consistency constraints under different data augmentations. Specifically, researchers impose consistency constraints on features between point clouds with different data augmentations or geometric calibrations, such as distinguishing points from different scenes through contrastive learning, constructing consistency constraints using color and geometric smoothness, or utilizing more complex consistency losses, such as JS-divergence and similarity-weighted losses. Given only sparse annotations, existing weakly supervised point cloud segmentation methods often fail to effectively learn scene context information due to the complex structure of point cloud data, resulting in poor segmentation performance. To explore the ability of existing methods to understand scene context, this embodiment validates them through mask evaluation: given instance-level partially masked point clouds as input, the model's segmentation performance is evaluated. (See attached...) Figure 4 As shown, even in this simple case, the performance of different data augmentation consistency methods drops sharply as the mask ratio increases, indicating their lack of effective understanding of the scene context.

[0005] Therefore, new methods are urgently needed to perform effective contextual understanding from point cloud data with complex structures and only a small number of annotations. Summary of the Invention

[0006] To address the problem of insufficient understanding of the context of complex point cloud data in existing methods, given only sparse annotations, this invention aims to provide a contextual point cloud modeling method and apparatus for weakly supervised point cloud segmentation, so as to effectively learn the scene context and obtain a point cloud semantic segmentation model with better performance.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a method for implementing contextual point cloud modeling for weakly supervised point cloud segmentation, comprising the following steps:

[0009] Obtain raw point cloud data, wherein the raw point cloud data package contains a small number of annotations;

[0010] The original point cloud data is augmented to obtain at least two augmented point cloud data sets.

[0011] Perform a masking operation on any of the enhanced point cloud data to obtain masked point cloud data;

[0012] A point cloud segmentation model is established, and point-by-point classification probabilities are generated based on the two enhanced point cloud data, the masked point cloud data, and the point cloud segmentation model.

[0013] A feature consistency loss is generated based on the point-by-point classification probabilities of the two enhanced point cloud data. A first cross-entropy loss and a second cross-entropy loss are generated based on the point-by-point classification probabilities of the two enhanced point cloud data respectively. A first mask feature consistency loss is generated based on the point-by-point classification probabilities of one enhanced point cloud data and the point-by-point classification probabilities of the masked point cloud data. A second mask feature consistency loss is generated based on the point-by-point classification probabilities of the other enhanced point cloud data and the point-by-point classification probabilities of the masked point cloud data.

[0014] A cross-entropy loss is generated based on the first cross-entropy loss and the second cross-entropy loss, and a mask feature consistency loss is generated based on the first mask feature consistency loss and the second mask feature consistency loss.

[0015] A training model for training point cloud data is generated based on the feature consistency loss, the cross-entropy loss, and the mask feature consistency loss.

[0016] Secondly, the present invention provides a contextual point cloud modeling implementation device for weakly supervised point cloud segmentation, comprising:

[0017] A data acquisition unit is used to acquire raw point cloud data, wherein the raw point cloud data package contains a small number of annotations;

[0018] The data processing unit is used to perform the following steps:

[0019] The original point cloud data is augmented to obtain at least two augmented point cloud data sets.

[0020] Perform a masking operation on any of the enhanced point cloud data to obtain masked point cloud data;

[0021] A point cloud segmentation model is established, and point-by-point classification probabilities are generated based on the two enhanced point cloud data, the masked point cloud data, and the point cloud segmentation model.

[0022] A feature consistency loss is generated based on the point-by-point classification probabilities of the two enhanced point cloud data. A first cross-entropy loss and a second cross-entropy loss are generated based on the point-by-point classification probabilities of the two enhanced point cloud data respectively. A first mask feature consistency loss is generated based on the point-by-point classification probabilities of one enhanced point cloud data and the point-by-point classification probabilities of the masked point cloud data. A second mask feature consistency loss is generated based on the point-by-point classification probabilities of the other enhanced point cloud data and the point-by-point classification probabilities of the masked point cloud data.

[0023] A cross-entropy loss is generated based on the first cross-entropy loss and the second cross-entropy loss, and a mask feature consistency loss is generated based on the first mask feature consistency loss and the second mask feature consistency loss.

[0024] A training model for training point cloud data is generated based on the feature consistency loss, the cross-entropy loss, and the mask feature consistency loss.

[0025] Thirdly, the present invention also provides an electronic device, including a processor and a memory;

[0026] The memory is used to store programs;

[0027] The processor executes the program to implement the method described above.

[0028] Fourthly, the present invention also provides a computer-readable storage medium storing a program that is executed by a processor to implement the methods described above.

[0029] Fifthly, the present invention also provides a computer program product or computer program comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions to cause the computer device to perform the preceding method.

[0030] Compared with the prior art, the beneficial effects of this invention are as follows: This invention provides a segmentation method for weakly labeled and structurally complex point cloud data, which allows the model to deepen its understanding of the context of the point cloud scene, effectively improves the segmentation performance of the point cloud segmentation model under weak labeling, and reduces the strong dependence of the point cloud segmentation model on dense labeling.

[0031] First, using the construction method proposed in this invention, continuous masking is performed on regional point cloud data to construct a mask feature prediction task that is beneficial for the model to learn the context of point cloud data. Second, the training method proposed in this invention is adopted. This method constrains the consistency of corresponding features before and after the masking of the same point cloud data, enabling the model to effectively utilize neighborhood information in the mask data, thereby improving the model's contextual understanding of point cloud data. Ultimately, this achieves semantic segmentation performance of the point cloud segmentation model in weakly labeled scenarios. This invention effectively solves the problem of point cloud semantic segmentation in scenarios with very sparse annotations and very complex point cloud structures. It can be widely applied to large-scale 3D visual understanding applications, such as autonomous driving and urban exploration, breaking the strong dependence of fully supervised point cloud segmentation methods on dense and expensive annotations, and greatly increasing the applicability of point cloud segmentation models. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of the overall training method for contextual point cloud modeling for weakly supervised point cloud segmentation according to an embodiment of the present invention;

[0034] Figure 2 This is a region-based point cloud masking method in this embodiment of the invention;

[0035] Figure 3 This is an example of the effect of the region-based point cloud masking method in this embodiment of the invention;

[0036] Figure 4 These are experimental results from embodiments of the present invention verifying the ability of the point cloud segmentation method to understand the contextual scene;

[0037] Figure 5 This is a schematic diagram of the device structure in an embodiment of the present invention.

[0038] Figure 6 This is a schematic diagram of an electronic device in an embodiment of the present invention. Detailed Implementation

[0039] 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 embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0040] Example:

[0041] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, in the embodiments of this invention are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0042] Example 1

[0043] To address the problem of insufficient understanding of the context of complex point cloud data in existing technologies, given only sparse annotations, this invention aims to provide a contextual point cloud modeling method for weakly supervised point cloud segmentation, so as to effectively learn the scene context and obtain a point cloud semantic segmentation model with better performance.

[0044] See Figures 1 to 4 A contextual point cloud modeling method for weakly supervised point cloud segmentation may specifically include the following steps:

[0045] Step 101: Obtain raw point cloud data, which contains a small number of annotations.

[0046] In this step, the raw point cloud data is acquired. The point cloud data consists of N points P = {p1, p2, ..., p...} N A set consisting of}, where each point p n It typically includes coordinate position and RGB information, i.e., p n =P[n]=(x n ,y n ,z n ,r n ,g n ,b nIt should be noted that the "small amount of annotation Y" mentioned in the embodiments of the present invention refers to an annotation quantity of ≤0.2%. However, in fact, the method provided by the embodiments of the present invention can also be used for cases with an extremely small amount of annotation Y, where an extremely small amount of annotation Y refers to an annotation quantity of ≤0.02%.

[0047] It should also be noted that in the embodiments of the present invention, [·] is used as an index operation to indicate the retrieval of the corresponding element (which can be a vector or scalar) from a set or matrix.

[0048] Step 102: Perform enhancement operations on the original point cloud data to obtain at least two enhanced point cloud data.

[0049] In this step, to learn from the large number of unlabeled points in the original point cloud data, two random augmentations are applied to obtain at least two augmented point cloud data sets P1 = Aug1(P) and P2 = Aug2(P). Specifically, one or more data augmentation methods can be used, such as randomly discarding some points, randomly flipping horizontally, randomly changing color contrast, randomly changing color levels, and randomly adding color noise. After these methods, for each point in P1 and P2, their features are different, but their semantics are the same (corresponding to the same category), which is used for subsequent point cloud feature extraction and feature consistency learning. Since no point cloud annotations are required, the model can learn from a large amount of unlabeled point cloud data through this process.

[0050] Step 103: Perform a masking operation on any of the augmented point cloud data to obtain masked point cloud data. Although the aforementioned feature consistency loss can be learned from a large amount of unlabeled point cloud data, the structure of point cloud data is often very complex. Simply utilizing the feature consistency under different data augmentations is insufficient to effectively learn contextual information and thus obtain better point cloud segmentation performance. To enable the model to learn contextual information, some contextual information is first removed from the point cloud data.

[0051] In this step, to reasonably remove some contextual information from the enhanced point cloud data, this embodiment of the invention introduces a region-based point cloud masking method. This method uniformly divides the scene into cubes and masks the RGB information of randomly selected points within the cubes. This embodiment of the invention uses the proposed region-based masking method to mask the enhanced point cloud data, obtaining masked point cloud data, which is then used for the masked feature reconstruction task.

[0052] Specifically, the above content can be divided into the following sub-steps:

[0053] Step 1031: Based on the original point cloud data and a region of a set size, obtain a series of cuboids for dividing the spatial data of the original point cloud data.

[0054] Step 1032: Based on the set mask ratio and the cuboid, a series of cuboid subsets to be masked are obtained.

[0055] Step 1033: Perform a masking operation on all the original point cloud data within the cuboid subset.

[0056] For example, given a region size G to represent the number of cubes, and cubes parallel to the axes in the 3D coordinate system are [(x min ,y min ,z min ),(x max ,y max ,z max The smallest cube representing the point cloud is [(0,0,0),(l,w,h)]. In this embodiment of the invention, the scene is uniformly divided into a set of cubic regions. Right now As shown below:

[0057]

[0058] Where, x i y j z k The points are evenly divided along the x, y, and z axes, (l G ,w G ,h G These represent the length, width, and height of the region, respectively. Then, in this embodiment, R·G is randomly selected. 3 each region And calculate the mask flag M as follows:

[0059]

[0060] Where ∈ indicates whether a point is inside a cube. Then, the masked point cloud P is calculated as follows: m :

[0061] P m [n] = [x] n ,y n ,z n ,M[n]·r n ,M[n]·g n ,M[n]·b n ],

[0062] Since sparse convolution operations in 3D UNet require coordinate information to construct the convolution kernel, the x, y, and z coordinates remain unchanged. (See attached image) Figure 3As shown in (a)-(c), compared to point-by-point masking methods, the region-based masking method proposed in this embodiment can continuously mask disordered and irregular point clouds, providing meaningful contextual patterns to be supplemented, such as partial internal instance masking and cross-instance masking. Furthermore, this method can flexibly handle different numbers of annotations by adjusting the region size.

[0063] Step 104: Establish a point cloud segmentation model, and generate point-by-point classification probabilities based on the two enhanced point cloud data, the masked point cloud data, and the point cloud segmentation model.

[0064] In this step, establishing the point cloud segmentation model involves constructing a point cloud segmentation network, given a segmentation model f parameterized by θ. θ (·) For example, 3D UNet, this model can generate point-by-point classification probabilities Z = Softmax(f θ (P)), where Z[n]∈(0,1), argmax(Z[n])∈C and {0,1,2,…,C-1} is a predefined set of categories with class C.

[0065] Step 105: Generate feature consistency loss based on the point-by-point classification probabilities of the two enhanced point cloud data sets; generate a first cross-entropy loss and a second cross-entropy loss based on the point-by-point classification probabilities of the two enhanced point cloud data sets respectively; generate a first mask feature consistency loss based on the point-by-point classification probabilities of one enhanced point cloud data set and the point-by-point classification probabilities of the masked point cloud data set; generate a second mask feature consistency loss based on the point-by-point classification probabilities of the other enhanced point cloud data set and the point-by-point classification probabilities of the masked point cloud data set; generate a cross-entropy loss based on the first cross-entropy loss and the second cross-entropy loss; and generate a mask feature consistency loss based on the first mask feature consistency loss and the second mask feature consistency loss.

[0066] In this step, the present invention combines the proposed context mask-based training method to train the cross-entropy loss function on a small number of labeled parts of the point cloud data after different data augmentation, to train the feature consistency of the point cloud data after different data augmentation by minimizing JS-divergence, and to train the feature consistency of the features before and after masking by minimizing mask feature consistency.

[0067] Repeat the above steps until the point cloud segmentation model converges to a better validation set metric.

[0068] Specifically, in order for the model to learn category information from point cloud data, the point cloud segmentation model Softmax(f) is first used. θ(·) Extract the corresponding point-by-point classification probabilities Z1 and Z2. Then, for the labeled points in the point cloud data, calculate the cross-entropy loss:

[0069]

[0070] Where CE(·,·) is the cross-entropy loss function, and the cross-entropy loss... It consists of the first cross-entropy loss and the second cross-entropy loss.

[0071] To enable the model to learn information from the numerous unlabeled points in point cloud data, constraint is placed on the consistency of point cloud features under different data augmentations. Feature consistency loss is used in this process. Constrain pointwise feature consistency in the following ways:

[0072]

[0073] JS(·,·) is used to minimize the Jensen-Shannon (JS) divergence between different features. Feature consistency from different enhancements can utilize unlabeled data, but it is still insufficient to understand the complex structure of point cloud data and cannot effectively explore key information such as spatial, color, and semantic continuity to achieve satisfactory segmentation results.

[0074] To enable the model to effectively learn contextual information from a large amount of unlabeled point cloud data, this embodiment of the invention learns contextual information by constraining the consistency between masked and unmasked features. To this end, this embodiment proposes minimizing the distribution gap between masked and unmasked features. In this way, the point cloud segmentation model will learn to utilize the unmasked portion of the masked point cloud, i.e., the surrounding context, thereby improving segmentation performance. Specifically, for the point-by-point classification probabilities Z1, Z2, Z3 extracted from two randomly augmented and masked point clouds respectively... m The embodiments of the present invention introduce the following mask feature consistency loss:

[0075]

[0076] Among them, the unmasked pointwise classification probabilities Z1 and Z2 are regarded as the masked pointwise classification probabilities Z. m The "target value". Meanwhile, when calculating the mask feature consistency loss, the gradients of the pointwise classification probabilities Z1 and Z2 of the unmasked data are broken, and the mask feature consistency loss... It consists of the first mask feature consistency loss and the second mask feature consistency loss.

[0077] Step 106: Generate a training model for training point cloud data based on the feature consistency loss, the cross-entropy loss, and the mask feature consistency loss.

[0078] In this step, in order to simultaneously learn the category knowledge in the point cloud data, the knowledge of a large number of unlabeled points in the point cloud data, and the contextual knowledge in the point cloud data, this embodiment of the invention weights the aforementioned cross-entropy loss, feature consistency loss, and mask-based feature consistency loss to obtain the overall training objective as follows:

[0079]

[0080] Where α is a hyperparameter controlling the strength of consistency loss optimization, and β is a hyperparameter controlling the strength of context mask learning optimization. For example... Figure 4 As shown, compared to methods that only consider cross-entropy loss and feature consistency loss, the context mask-based training method proposed in this embodiment can still obtain a relatively accurate mIoU index when the point cloud occlusion probability is high. This indicates that the method in this embodiment can effectively learn the context in point cloud data and improve the performance of point cloud segmentation.

[0081] Supervised learning is performed on point cloud data with a small number of labels to enable the model to have a preliminary ability to recognize different objects in the scene; unsupervised learning based on feature consistency and mask feature consistency is performed on a large number of unlabeled data to improve the model's generalization ability and context modeling ability for different scenes.

[0082] By applying different strengths of label-based cross-entropy loss function constraints, JS-divergence minimization constraints based on different data augmentations, and JS-divergence constraints based on point cloud features before and after masking to the three different features mentioned above, the model can effectively utilize a small amount of labeled information and a large amount of unlabeled information during training to obtain the target model.

[0083] During the testing or inference phase, the trained point cloud segmentation model directly classifies the user-input point cloud data point by point to output the point cloud segmentation results without any point cloud annotation.

[0084] Example 2

[0085] The technical solutions in the embodiments of the present invention will be described below with reference to actual data:

[0086] This embodiment of the invention applies the method described above to two commonly used point cloud segmentation datasets: ScanNet V2, which has 20 semantic categories, with 1,201 training samples, 312 validation samples, and 100 test samples. The embodiment evaluates the point cloud semantic segmentation model on the validation set and online test set using the previous method. S3DIS is a large-scale point cloud dataset containing 6 regions, 271 rooms, and 13 semantic categories. This embodiment uses the widely used area5 test set to evaluate the model's performance, with 204 training scans and 68 test scans. The scene segmentation results on the ScanNet V2 and S3DIS datasets are shown in Table 1 and Table 2 below, respectively.

[0087] Table 1 shows the experimental results on the S3DIS dataset in Example 2.

[0088]

[0089] Table 2 shows the experimental results on the ScanNetV2 dataset in Example 2.

[0090]

[0091] Example 3

[0092] See Figure 5 Based on the same inventive concept, embodiments of the present invention also provide a contextual point cloud modeling implementation device for weakly supervised point cloud segmentation, comprising:

[0093] A data acquisition unit is used to acquire raw point cloud data, wherein the raw point cloud data package contains a small number of annotations;

[0094] The data processing unit is used to perform the following steps:

[0095] The original point cloud data is augmented to obtain at least two augmented point cloud data sets.

[0096] Perform a masking operation on any of the enhanced point cloud data to obtain masked point cloud data;

[0097] A point cloud segmentation model is established, and point-by-point classification probabilities are generated based on the two enhanced point cloud data, the masked point cloud data, and the point cloud segmentation model.

[0098] A feature consistency loss is generated based on the point-by-point classification probabilities of the two enhanced point cloud data. A first cross-entropy loss and a second cross-entropy loss are generated based on the point-by-point classification probabilities of the two enhanced point cloud data respectively. A first mask feature consistency loss is generated based on the point-by-point classification probabilities of one enhanced point cloud data and the point-by-point classification probabilities of the masked point cloud data. A second mask feature consistency loss is generated based on the point-by-point classification probabilities of the other enhanced point cloud data and the point-by-point classification probabilities of the masked point cloud data.

[0099] A cross-entropy loss is generated based on the first cross-entropy loss and the second cross-entropy loss, and a mask feature consistency loss is generated based on the first mask feature consistency loss and the second mask feature consistency loss.

[0100] A training model for training point cloud data is generated based on the feature consistency loss, the cross-entropy loss, and the mask feature consistency loss.

[0101] Since this system is the system corresponding to the blockchain-based weakly supervised point cloud segmentation context point cloud modeling implementation method of this invention, and the principle of solving the problem in this system is similar to that of this method, the implementation of this system can refer to the implementation process of the above method embodiment, and the repeated parts will not be described again.

Claims

1. A method for implementing contextual point cloud modeling for weakly supervised point cloud segmentation, characterized in that, Including the following steps: Obtain raw point cloud data, wherein the raw point cloud data package contains annotations; The original point cloud data is augmented to obtain at least two augmented point cloud data sets. Perform a masking operation on any of the enhanced point cloud data to obtain masked point cloud data; A point cloud segmentation model is established, and point-by-point classification probabilities are generated based on the two enhanced point cloud data, the masked point cloud data, and the point cloud segmentation model. A feature consistency loss is generated based on the point-by-point classification probabilities of the two enhanced point cloud data. A first cross-entropy loss and a second cross-entropy loss are generated based on the point-by-point classification probabilities of the two enhanced point cloud data respectively. A first mask feature consistency loss is generated based on the point-by-point classification probabilities of one enhanced point cloud data and the point-by-point classification probabilities of the masked point cloud data. A second mask feature consistency loss is generated based on the point-by-point classification probabilities of the other enhanced point cloud data and the point-by-point classification probabilities of the masked point cloud data. A cross-entropy loss is generated based on the first cross-entropy loss and the second cross-entropy loss, and a mask feature consistency loss is generated based on the first mask feature consistency loss and the second mask feature consistency loss. A training model for training point cloud data is generated based on the feature consistency loss, the cross-entropy loss, and the mask feature consistency loss. Specifically, based on the original point cloud data and a region of a set size, a series of cuboids are obtained to divide the spatial data of the original point cloud data; based on the set masking ratio and the cuboids, a series of cuboid subsets to be masked are obtained; and a masking operation is performed on all the original point cloud data within the cuboid subsets. The masked point cloud P is calculated using the following method. m : P m [n]=[x n ,y n ,z n ,M[n]·r n ,M[n]·g n ,M[n]·b n ]; In the formula, x n ,y n ,z n For point p n The coordinate position, r n ,g n ,b n For point p n The RGB information, M[n] is the mask flag bit, ∈ indicates whether a point is inside a cube. The target area to be masked.

2. The method for implementing contextual point cloud modeling for weakly supervised point cloud segmentation according to claim 1, characterized in that, The cross-entropy loss Specifically: In the formula, CE(·,·) represents the cross-entropy loss function. Z1 and Z2 represent the point-by-point classification probabilities of the two augmented point cloud data, respectively. Y represents the label, and the subscript s indicates the labeled point in the label index set.

3. The method for implementing contextual point cloud modeling for weakly supervised point cloud segmentation according to claim 2, characterized in that, The feature consistency loss Specifically: In the formula, JS(·,·) represents minimizing the divergence function between different features, Z1 and Z2 are the point-by-point classification probabilities of the two enhanced point cloud data, n represents the nth point cloud data, and N represents the set of a certain point cloud data.

4. The method for implementing contextual point cloud modeling for weakly supervised point cloud segmentation according to claim 3, characterized in that, The mask feature consistency loss Specifically: In the formula, JS(·,·) represents minimizing the divergence function between different features, n represents the nth point cloud data, and Z1, Z2, Z... m These represent the point-by-point classification probabilities of the two enhanced point cloud datasets and the point-by-point classification probabilities of the masked point cloud dataset, respectively. n represents the nth point cloud dataset, and N represents the set of a certain point cloud dataset.

5. The method for implementing contextual point cloud modeling for weakly supervised point cloud segmentation according to claim 4, characterized in that, The overall training objective of the training model used to train point cloud data. In the formula, α is a hyperparameter that controls the optimization strength of consistency loss, and β is a hyperparameter that controls the optimization strength of context mask learning.

6. A device for implementing contextual point cloud modeling for weakly supervised point cloud segmentation, characterized in that, include: A data acquisition unit is used to acquire raw point cloud data, wherein the raw point cloud data package contains a small number of annotations; The data processing unit is used to perform the following steps: The original point cloud data is augmented to obtain at least two augmented point cloud data sets. Perform a masking operation on any of the enhanced point cloud data to obtain masked point cloud data; A point cloud segmentation model is established, and point-by-point classification probabilities are generated based on the two enhanced point cloud data, the masked point cloud data, and the point cloud segmentation model. A feature consistency loss is generated based on the point-by-point classification probabilities of the two enhanced point cloud data. A first cross-entropy loss and a second cross-entropy loss are generated based on the point-by-point classification probabilities of the two enhanced point cloud data respectively. A first mask feature consistency loss is generated based on the point-by-point classification probabilities of one enhanced point cloud data and the point-by-point classification probabilities of the masked point cloud data. A second mask feature consistency loss is generated based on the point-by-point classification probabilities of the other enhanced point cloud data and the point-by-point classification probabilities of the masked point cloud data. A cross-entropy loss is generated based on the first cross-entropy loss and the second cross-entropy loss, and a mask feature consistency loss is generated based on the first mask feature consistency loss and the second mask feature consistency loss. A training model for training point cloud data is generated based on the feature consistency loss, the cross-entropy loss, and the mask feature consistency loss. Specifically, based on the original point cloud data and a region of a set size, a series of cuboids are obtained to divide the spatial data of the original point cloud data; based on the set masking ratio and the cuboids, a series of cuboid subsets to be masked are obtained; and a masking operation is performed on all the original point cloud data within the cuboid subsets. The masked point cloud P is calculated using the following method. m : P m [n]=[x n ,y n ,z n ,M[n]·r n ,M[n]·g n ,M[n]·b n ]; In the formula, x n ,y n ,z n For point p n The coordinate position, r n ,g n ,b n For point p n The RGB information, M[n] is the mask flag bit, ∈ indicates whether a point is inside a cube. The target area to be masked.

7. An electronic device, characterized in that, The electronic device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set, or an instruction set, and the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the context point cloud modeling implementation method for weakly supervised point cloud segmentation as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or instruction set is loaded and executed by a processor to implement the context point cloud modeling implementation method for weakly supervised point cloud segmentation as described in any one of claims 1 to 5.

Images