Weakly supervised point cloud splicing data augmentation implementation method and device and medium

Abstract

The application discloses a weakly supervised point cloud splicing data enhancement implementation method and device and medium, and relates to the technical field of data enhancement in point cloud segmentation. The method comprises the following steps: acquiring original point cloud scene data, wherein the original point cloud scene contains a small amount of labels; performing non-overlapping point cloud splicing data enhancement on the original point cloud scene data to obtain point cloud scene enhancement data fused with multiple point cloud scene information; establishing a point cloud segmentation model, inputting the point cloud scene enhancement data into the point cloud segmentation model to generate point-by-point classification probability; and training the point cloud segmentation model according to the point-by-point classification probability and cross-entropy loss. The application effectively solves the point cloud semantic segmentation problem in a scene where the labels are very sparse and the point cloud structure is very complex, and can be widely used in large-scale 3D visual understanding application scenarios.

Description

Technical Field

[0001] This invention relates to the technical field of data augmentation in point cloud segmentation, and in particular to a method, apparatus and medium for data augmentation of weakly supervised point cloud stitching. Background Technology

[0002] With the widespread adoption of inexpensive LiDAR sensors and depth cameras, scene understanding based on 3D environments, including point cloud semantic segmentation, object detection, and instance segmentation, has become fundamental to autonomous driving and robot navigation, and a hot research topic in the computer vision field. Among the many 3D scene understanding tasks, point cloud semantic segmentation is one of the most basic, providing fine-grained scene information through point-by-point classification. To achieve satisfactory performance, researchers have designed and trained 3D segmentation models using densely annotated datasets that provide point-by-point labels; however, this is very time-consuming and involves dense labeling.

[0003] In recent years, to reduce reliance on dense annotations, semantic segmentation research has primarily focused on learning from sparse labels. The core challenge lies in learning from data consisting of a very limited number of labeled parts and a large proportion of unlabeled parts. For the labeled parts, standard cross-entropy loss can be directly applied. The difficulty lies in learning from the unlabeled parts. A straightforward approach is to use a model trained on a limited number of labels to predict pseudo-labels for unlabeled points to further train the model. However, models trained on a small subset of data are prone to overfitting and outputting inaccurate pseudo-labels. To avoid this risk, existing methods typically employ multi-stage or multi-branch approaches to enforce feature consistency between two different augmented (or perturbed) versions of the same point cloud data. Multi-stage methods are simple but time-consuming, while multi-branch methods inject many components (such as point cloud augmenters, student-teacher models, and consistency loss designs) into fully supervised methods, making the entire framework difficult to implement, analyze, and most importantly, train. This situation raises the question: Is there a way to fully leverage fully supervised methods and introduce only minimal modifications to make them perform well in scenarios with limited annotations?

[0004] Existing methods can be seen as a derivative of semi-supervised learning in image classification scenarios. Semi-supervised learning is one of the most thoroughly researched areas for handling partially labeled data (i.e., only a few images have image-level labels). Therefore, since information cannot be easily obtained from unlabeled images, multi-branch training methods are needed. However, in the case of weakly supervised point cloud segmentation, the model struggles to learn complete contextual information from limited annotations, and multi-branch training methods are extremely costly. The existing Mix3D method alleviates the problem of model overfitting to a small amount of labeled data by directly fusing point clouds to obtain enhanced point cloud data. However, this method directly overlaps and mixes point clouds, damaging the original scene information and hindering the model's learning of contextual information. Therefore, effective learning from structurally complex point cloud data with only a few annotations urgently requires new methods. Summary of the Invention

[0005] In order to at least partially solve one of the technical problems existing in the prior art, the purpose of this invention is to provide a method, apparatus and medium for enhancing weakly supervised point cloud stitching data.

[0006] The first technical solution adopted in this invention is:

[0007] A method for augmenting weakly supervised point cloud stitching data includes the following steps:

[0008] Obtain raw point cloud scene data, wherein the raw point cloud scene contains a small number of annotations;

[0009] The original point cloud scene data is enhanced by non-overlapping point cloud stitching data to obtain point cloud scene enhancement data that integrates information from multiple point cloud scenes;

[0010] A point cloud segmentation model is established, and the enhanced point cloud scene data is input into the point cloud segmentation model to generate point-by-point classification probabilities.

[0011] The point cloud segmentation model is trained based on the point-by-point classification probability and cross-entropy loss.

[0012] Furthermore, the point cloud scene data consists of N points P = {p1, p2, ..., p...} N A set consisting of}, where each point p n Includes coordinate position and RGB information, i.e., p n =P[n]=(x n ,y n ,z n ,r n ,g n ,b n ).

[0013] Furthermore, the step of performing non-overlapping point cloud stitching data enhancement on the original point cloud scene data to obtain point cloud scene enhancement data that integrates information from multiple point cloud scenes includes:

[0014] Multiple point cloud scene data are randomly sampled from the original point cloud scene data to form a sampled point cloud scene data set;

[0015] Perform coordinate normalization on each point cloud scene data in the sampled point cloud scene data set;

[0016] Calculate the shape information of the length of the sampled point cloud scene on the x-axis, y-axis, and z-axis, respectively;

[0017] One sampling point cloud scene is randomly selected from the sampling point cloud scene data set as the baseline point cloud scene data set, and the other sampling point cloud scene data sets are used as non-baseline sampling point cloud scene data sets.

[0018] The non-reference sampled point cloud scene data set is traversed, and the coordinate data of the non-reference sampled point cloud scene is adjusted according to the shape information of the sampled point cloud scene. The point cloud scene data is then stitched together without overlap to form point cloud scene enhancement data.

[0019] Further, the step of traversing the non-reference sampled point cloud scene data set, adjusting the coordinate data of the non-reference sampled point cloud scene according to the shape information of the sampled point cloud scene, and stitching the point cloud scene data non-overlappingly to form point cloud scene enhancement data includes:

[0020] A1. Randomly select a non-reference sampling point cloud scene data P from the non-reference sampling point cloud scene coordinate data set. i And remove that data from the set;

[0021] A2. Based on the position information of the non-reference sampling point cloud scene data and the position information of the reference point cloud scene data, calculate the minimum offset on each axis that ensures that the non-reference sampling point cloud scene data does not overlap with the reference point cloud scene data.

[0022] A3. Randomly select an offset direction from the x-axis, y-axis, and z-axis, retain the offset value along that axis, and set the offset values ​​of the other two axes to 0. Update the non-referenced sampling point cloud scene data according to the offset value to obtain the updated non-referenced sampling point cloud scene data P. i ′;

[0023] A4. Combine the baseline sampled point cloud scene data with the updated non-baseline sampled point cloud scene data P i The data are stitched together to obtain the updated baseline sampled point cloud scene data, and the position information of the baseline sampled point cloud scene data is updated according to the stitched coordinate data.

[0024] A4. Repeat steps A1 to A4 until the non-baseline sampling point cloud scene data set is empty.

[0025] Furthermore, the expression for the point cloud segmentation model is:

[0026] Z = Softmax(f θ (P))

[0027] In the formula, Z represents the point-by-point classification probability of the point cloud scene augmented data, and f θ (·) represents the point cloud segmentation model parameterized by θ, Softmax(·) is the normalized exponential function, and P is the point cloud scene data.

[0028] Furthermore, the cross-entropy loss The expression is:

[0029]

[0030] In the formula, CE(·,·) represents the cross-entropy loss function. Z represents the set of labeled indices, Z is the point-by-point classification probability of the point cloud scene augmentation data, Y is the label, and s represents the labeled points in the set of labeled indices.

[0031] Furthermore, the overall training objective used to train the point cloud segmentation model

[0032]

[0033] In the formula, α is a hyperparameter that controls the intensity of supervised loss optimization.

[0034] The second technical solution adopted in this invention is:

[0035] A device for enhancing weakly supervised point cloud stitching data includes:

[0036] Obtain raw point cloud scene data, wherein the raw point cloud scene contains a small number of annotations;

[0037] The original point cloud scene data is enhanced by non-overlapping point cloud stitching data to obtain point cloud scene enhancement data that integrates information from multiple point cloud scenes;

[0038] A point cloud segmentation model is established, and the enhanced point cloud scene data is input into the point cloud segmentation model to generate point-by-point classification probabilities.

[0039] The point cloud segmentation model is trained based on the point-by-point classification probability and cross-entropy loss.

[0040] The third technical solution adopted in this invention is:

[0041] A device for enhancing weakly supervised point cloud stitching data includes:

[0042] At least one processor;

[0043] At least one memory for storing at least one program;

[0044] When the at least one program is executed by the at least one processor, the at least one processor implements the method described above.

[0045] The fourth technical solution adopted in this invention is:

[0046] A computer-readable storage medium storing a processor-executable program, which, when executed by a processor, performs the method described above.

[0047] 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. By splicing and fusing multiple point cloud scene data, the contextual information of the scene is enhanced and supplemented, allowing the model to deepen its understanding of the context of the point cloud scene. This 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.

[0048] First, using the construction method proposed in this invention, multiple point cloud scene data are stitched together non-overlappingly, constructing new contextual information to supplement the original contextual information without losing the original scene information. Second, the point cloud segmentation model is trained using cross-entropy loss, achieving excellent semantic segmentation performance in weakly labeled scenarios. This invention effectively solves the point cloud semantic segmentation problem in scenarios with very sparse annotations and very complex point cloud structures. Moreover, it can also improve model performance in fully supervised scenarios, and can be widely applied to large-scale 3D visual understanding applications, such as autonomous driving and urban exploration. It breaks the strong dependence of fully supervised point cloud segmentation methods on dense and expensive annotations, greatly increasing the applicability of the point cloud segmentation model. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following description is provided with accompanying drawings of the relevant technical solutions in the embodiments of the present invention or the prior art. It should be understood that the accompanying drawings described below are only for the purpose of clearly illustrating some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1This is a flowchart illustrating the steps of a method for enhancing weakly supervised point cloud stitching data in an embodiment of the present invention.

[0051] Figure 2 This is a flowchart of the point cloud stitching data enhancement method in an embodiment of the present invention. Detailed Implementation

[0052] The embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0053] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0054] In the description of this invention, "several" means one or more, "multiple" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. If "first" or "second" is used, it is only for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features. Furthermore, "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0055] In the description of this invention, unless otherwise explicitly defined, terms such as "setting," "installing," and "connecting" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0056] To address the issues of existing technologies focusing on local details while ignoring contextual information and the high overhead of multi-branch training, given only sparse annotations, this invention aims to provide a method for augmenting weakly supervised point cloud stitching data, supplementing and learning scene context to obtain a point cloud semantic segmentation model with better performance.

[0057] like Figure 1 As shown in the figure, this embodiment provides a method for augmenting weakly supervised point cloud stitching data, which may specifically include the following steps:

[0058] Step S1: Obtain the original point cloud scene data, which contains a small number of annotations.

[0059] In this step, the original point cloud scene data is acquired. The point cloud scene 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 n It should be noted that the "small amount of annotations 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 annotations Y, which refers to an annotation quantity of ≤0.02%. The embodiments of the present invention are still effective in a fully supervised scenario where Y=100%.

[0060] 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.

[0061] Step S2: Perform point cloud stitching data enhancement operation on the original point cloud scene data to obtain enhanced point cloud scene data that integrates multiple point cloud scenes.

[0062] See Figure 2 In some embodiments, step S2 can be divided into the following sub-steps:

[0063] Step S21: Randomly sample M point cloud scene data from the original point cloud scene data to obtain a sampled point cloud scene data set.

[0064] Step S22: Normalize the coordinates of each point cloud scene data in the sampled point cloud scene data set. Through normalization, the origin of the coordinates is placed at the center of the point cloud scene.

[0065] Step S23: Calculate the shape information of the length of the sampled point cloud scene on the x-axis, y-axis, and z-axis. l = max(x) - min(x), w = max(y) - min(y), h = max(z) - min(z), where max(·) and min(·) represent the operations of finding the maximum and minimum values ​​of the corresponding coordinates in a point cloud scene, respectively. x, y, and z represent the coordinate values ​​of the point cloud scene data, respectively.

[0066] Step S24: Randomly select a point cloud scene from the sampled point cloud scene data set as the reference point cloud scene data. Other sampling point cloud scene data As a collection of cloud scene data from non-benchmark sampling points.

[0067] Step S25: Traverse the non-reference sampling point cloud scene data set, adjust the non-reference sampling point cloud scene coordinate data, and stitch the point cloud scene data together without overlap to form the enhanced point cloud scene data.

[0068] As an optional implementation, step S25 can be divided into the following sub-steps:

[0069] Step S251: Randomly select a non-reference sampling point cloud scene data from the non-reference sampling point cloud scene coordinate data set. Then remove that data from the set.

[0070] Step S252: Based on the location information of the non-reference sampling point cloud scene data... i ,w i ,h i and the location information of the reference point cloud scene data base ,w base ,h base Calculate the minimum offset Δx = |l on each axis that ensures the non-reference sampled point cloud scene data does not overlap with the reference point cloud scene data. base -l i |,Δy=|w base -w i |,Δz=|h base -h i |

[0071] Step S253: Randomly select an offset direction from the x-axis, y-axis, and z-axis, retain the offset value along that axis, and set the offset values ​​of the other two axes to 0. Update the non-reference sampling point cloud scene data x = x + Δx, y = y + Δy, z = z + Δz according to the offset values ​​to obtain the updated non-reference sampling point cloud scene data P. i ′.

[0072] Step S254: Combine the baseline sampled point cloud scene data with the updated non-baseline sampled point cloud scene data P i By stitching them together, we obtain the updated baseline sampled point cloud scene data P. base At the same time, the location information of the benchmark sampling point cloud scene data is updated based on the stitched coordinate data.

[0073] Step S255: Repeat steps S251 to S254 until the non-reference sampling point cloud scene data set is empty.

[0074] Step S3: Establish a point cloud segmentation model, and generate point-by-point classification probabilities based on the enhanced point cloud data and the point cloud segmentation model.

[0075] 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 It is a predefined set of categories with class C; in order to learn the category information in the original point cloud data, cross-entropy loss needs to be calculated for the labeled points:

[0076]

[0077] In the formula, CE(·,·) represents the cross-entropy loss function. Z represents the set of labeled indices, Z is the point-by-point classification probability of the augmented point cloud data, Y is the label, and the subscript s indicates the labeled point in the set of labeled indices.

[0078] Step S4: Train the point cloud segmentation model based on the point-by-point classification probability and cross-entropy loss.

[0079] In this step, in order to learn category knowledge from point cloud data simultaneously, the overall training objective is as follows:

[0080]

[0081] Where α is a hyperparameter that controls the intensity of loss optimization.

[0082] 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.

[0083] The technical solutions in the embodiments of the present invention will be described below with reference to actual data:

[0084] 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.

[0085] Table 1 shows the experimental results on the S3DIS dataset in this embodiment.

[0086] MinkNet 0.01％ 48.07 MinkNet+Mix3D 0.01％ 48.87 MinkNet+Shift (This article) 0.01％ 51.79 MinkNet 0.1％ 60.32 MinkNet+Mix3D 0.1％ 60.44 MinkNet+Shift (This article) 0.1％ 61.64 MinkNet 100％ 63.8 MinkNet+Mix3D 100％ 62.98 MinkNet+Shift (This article) 100％ 64.58

[0087] Table 2 shows the experimental results on the ScanNetV2 dataset in this embodiment.

[0088]

[0089]

[0090] This embodiment also provides a device for enhancing weakly supervised point cloud stitching data, including:

[0091] Obtain raw point cloud scene data, wherein the raw point cloud scene contains a small number of annotations;

[0092] The original point cloud scene data is enhanced by non-overlapping point cloud stitching data to obtain point cloud scene enhancement data that integrates information from multiple point cloud scenes;

[0093] A point cloud segmentation model is established, and the enhanced point cloud scene data is input into the point cloud segmentation model to generate point-by-point classification probabilities.

[0094] The point cloud segmentation model is trained based on the point-by-point classification probability and cross-entropy loss.

[0095] This embodiment provides a device for enhancing weakly supervised point cloud stitching data. It can execute a method for enhancing weakly supervised point cloud stitching data provided in the method embodiment of the present invention, and can execute any combination of the implementation steps of the method embodiment. It has the corresponding functions and beneficial effects of the method.

[0096] This embodiment also provides a device for enhancing weakly supervised point cloud stitching data, including:

[0097] At least one processor;

[0098] At least one memory for storing at least one program;

[0099] When the at least one program is executed by the at least one processor, the at least one processor implements Figure 1 The method shown.

[0100] This embodiment provides a device for enhancing weakly supervised point cloud stitching data. It can execute a method for enhancing weakly supervised point cloud stitching data provided in the method embodiment of the present invention, and can execute any combination of the implementation steps of the method embodiment. It has the corresponding functions and beneficial effects of the method.

[0101] This application also discloses a computer program product or computer program, which includes 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, causing the computer device to perform... Figure 1 The method shown.

[0102] This embodiment also provides a storage medium storing instructions or programs that can execute the method embodiment of the present invention for enhancing weakly supervised point cloud stitching data. When the instructions or programs are run, any combination of implementation steps of the method embodiment can be executed, and the method has the corresponding functions and beneficial effects.

[0103] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is altered and sub-operations described as part of a larger operation are executed independently.

[0104] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0105] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0107] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0108] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0109] In the foregoing description of this specification, references to terms such as "one embodiment," "another embodiment," or "some embodiments" indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0110] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0111] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A weakly supervised point cloud stitching data augmentation oriented implementation method, characterized in that, Includes the following steps: Obtain raw point cloud scene data, wherein the raw point cloud scene contains a small number of annotations; The original point cloud scene data is enhanced by non-overlapping point cloud stitching data to obtain point cloud scene enhancement data that integrates information from multiple point cloud scenes; A point cloud segmentation model is established, and the enhanced point cloud scene data is input into the point cloud segmentation model to generate point-by-point classification probabilities. The point cloud segmentation model is trained based on the point-by-point classification probability and cross-entropy loss. The step of performing non-overlapping point cloud stitching data enhancement on the original point cloud scene data to obtain point cloud scene enhancement data that integrates information from multiple point cloud scenes includes: Multiple point cloud scene data are randomly sampled from the original point cloud scene data to form a sampled point cloud scene data set; Perform coordinate normalization on each point cloud scene data in the sampled point cloud scene data set; Calculate the shape information of the length of the sampled point cloud scene on the x-axis, y-axis, and z-axis, respectively; One sampling point cloud scene is randomly selected from the sampling point cloud scene data set as the baseline point cloud scene data set, and the other sampling point cloud scene data sets are used as non-baseline sampling point cloud scene data sets. Traverse the non-reference sampling point cloud scene data set, adjust the coordinate data of the non-reference sampling point cloud scene according to the shape information of the sampling point cloud scene, and stitch the point cloud scene data without overlap to form point cloud scene enhancement data; The process of traversing the non-reference sampled point cloud scene data set, adjusting the coordinate data of the non-reference sampled point cloud scene according to the shape information of the sampled point cloud scene, and stitching the point cloud scene data non-overlappingly to form point cloud scene enhancement data includes: A1. Randomly select a non-reference sampling point cloud scene data from the non-reference sampling point cloud scene coordinate data set. And remove that data from the set; A2. Based on the position information of the non-reference sampling point cloud scene data and the position information of the reference point cloud scene data, calculate the minimum offset on each axis that ensures that the non-reference sampling point cloud scene data does not overlap with the reference point cloud scene data. A3. Randomly select an offset direction from the x-axis, y-axis, and z-axis, retain the offset value along that axis, and set the offset values ​​of the other two axes to 0. Update the non-referenced sampling point cloud scene data according to the offset value to obtain the updated non-referenced sampling point cloud scene data. ; A4. Combine the baseline sampled point cloud scene data with the updated non-baseline sampled point cloud scene data. The data is stitched together to obtain the updated baseline sampled point cloud scene data, and the position information of the baseline sampled point cloud scene data is updated according to the stitched coordinate data. A4. Repeat steps A1 to A4 until the non-baseline sampling point cloud scene data set is empty.

2. The method for enhancing weakly supervised point cloud stitching data according to claim 1, characterized in that, Point cloud scene data is generated by Points The set of points Including coordinate position and RGB information, i.e. .

3. The method for enhancing weakly supervised point cloud stitching data according to claim 1, characterized in that, The expression for the point cloud segmentation model is: In the formula, To enhance the point-by-point classification probability of point cloud data, For the reason Parametric point cloud segmentation model, For normalized exponential functions, This is point cloud scene data.

4. The method for enhancing weakly supervised point cloud stitching data according to claim 1, characterized in that, The cross-entropy loss The expression is: In the formula, Represents the cross-entropy loss function. Represents the set of indices of the tags. To enhance the point-by-point classification probability of point cloud data, For annotation, This represents the labeled points within the set of indexes of the markers.

5. The method for enhancing weakly supervised point cloud stitching data according to claim 4, characterized in that, The overall training objective for training the point cloud segmentation model : In the formula, It is a hyperparameter that controls the intensity of supervised loss optimization.

6. A device for enhancing weakly supervised point cloud stitching data, used to implement the method described in any one of claims 1-5, characterized in that, include: Obtain raw point cloud scene data, wherein the raw point cloud scene contains a small number of annotations; The original point cloud scene data is enhanced by non-overlapping point cloud stitching data to obtain point cloud scene enhancement data that integrates information from multiple point cloud scenes; A point cloud segmentation model is established, and the enhanced point cloud scene data is input into the point cloud segmentation model to generate point-by-point classification probabilities. The point cloud segmentation model is trained based on the point-by-point classification probability and cross-entropy loss.

7. A device for enhancing weakly supervised point cloud stitching data, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method of any one of claims 1-5.

8. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the method as described in any one of claims 1-5.

Images