Active learning training method and system for three-dimensional point cloud segmentation model, and storage medium

Abstract

The invention discloses an active learning training method and system for a three-dimensional point cloud segmentation model, and a storage medium. The method comprises the following steps: providing a basic segmentation model, a basic probability model and a training data set; performing supervised pre-training to obtain a pre-training model; extracting embedded features for fitting training to obtain a joint probability model; calculating prototype characteristics; interpolation fusion is carried out to obtain fusion features, and joint probability representation is generated; calculating a stability score of the unlabeled data based on the fluctuation degree; performing redundancy removal and stability screening on the unlabeled data to obtain to-be-labeled data; receiving a marking operation on the to-be-marked data to form new marked data; performing semi-supervised training on the pre-training model by using the updated training data set to obtain an updated segmentation model; and iteration training is completed. According to the method, high-value samples are screened for labeling by constructing a joint probability modeling mechanism, and the discrimination capability and generalization performance of the model can be efficiently improved under limited labeling budget.

Description

[0001] This application is based on and claims priority to Chinese patent application No. 202511031359X, filed on July 23, 2025, entitled “Active Training Method, System, Medium and Device for 3D Point Cloud Segmentation Model”. Technical Field

[0002] This invention relates to the fields of computer vision and deep learning technology, and in particular to an active learning training method, system and storage medium for a 3D point cloud segmentation model. Background Technology

[0003] With the widespread application of 3D sensors such as LiDAR and RGB-D cameras, acquiring large-scale 3D point cloud data has become more convenient. As an important form of digital modeling, point cloud technology has been widely used in fields such as autonomous driving, intelligent robots, and augmented reality. In recent years, deep learning-based point cloud processing methods have made significant progress. However, these methods generally rely on a large amount of high-quality manually labeled data, requiring each point to be labeled individually, resulting in a time-consuming and costly labeling process. In addition, point cloud data is unstructured, sparse, and has large density variations, exhibiting significant differences in different application scenarios. Trained models are difficult to directly transfer to new scenarios, further exacerbating the demand for newly labeled data. Therefore, how to reduce the amount of labeling work while ensuring segmentation performance has become an urgent problem to be solved in the current field of point cloud learning.

[0004] To alleviate the pressure of labeling, existing research has proposed various efficient labeling learning methods, such as weakly supervised, semi-supervised, and small-sample methods. These methods typically use random sampling for labeling, failing to fully consider the uneven class distribution and density of point clouds, which can easily lead to missing key points or redundant labeling, thus limiting model performance improvement. Active learning, as an effective sample selection strategy, actively selects the most representative or uncertain samples for labeling by evaluating the information value of unlabeled samples, maximizing model performance with a limited labeling budget. However, existing active learning methods mostly rely on the softmax confidence of the model output as the sampling basis. Because deep neural networks are prone to "overconfidence," they can give high confidence even when predictions are wrong, leading to unreliable sample selection. In point cloud segmentation, due to blurred boundaries and complex structures, this problem is particularly prominent, affecting sampling accuracy and training efficiency.

[0005] To improve the reliability of uncertainty assessment, some studies have attempted to introduce probabilistic modeling. By constructing the distribution of features or prediction results and combining it with the spatial structure information of point clouds, these methods enhance the model's ability to express sample uncertainty. Such methods can more stably identify high-value samples, thereby improving the efficiency and accuracy of active learning. However, they still do not fully utilize the feature distribution of data to effectively extract key samples. Therefore, there is an urgent need for an active learning training method based on probabilistic modeling that can effectively utilize feature distribution information to guide the selection and annotation of key samples in 3D point clouds, thereby reducing annotation costs while improving the performance of segmentation models. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide an active learning training method, system, and storage medium for a three-dimensional point cloud segmentation model.

[0007] To achieve the aforementioned objectives, the technical solution adopted by this invention includes: In a first aspect, the present invention provides an active learning training method for a three-dimensional point cloud segmentation model, comprising: S1. Provide a basic segmentation model, a basic probability model, and a training dataset, wherein the training dataset includes labeled data and unlabeled data, and the true labels corresponding to the labeled data; S2. Supervised pre-training of the basic segmentation model is performed using the labeled data to obtain a pre-trained model; S3. Use the pre-trained model to perform forward inference on the training dataset, extract the embedding features, and input the embedding features into the basic probability model for fitting training to obtain the joint probability model; S4. Calculate the prototype features of each segmentation category based on the output features of the pre-trained model on the labeled data; interpolate and fuse the output features of the unlabeled data with the prototype features to obtain fused features; input the fused features into the joint probability model to generate a joint probability representation; calculate the stability score of the unlabeled data based on the fluctuation of the joint probability representation. S5. Redundancy removal and stability screening are performed on the unlabeled data to obtain the data to be labeled, wherein the stability screening selects a portion of the unlabeled data with lower stability scores; the labeling operation on the data to be labeled is received to form new labeled data, and the training dataset is updated; S6. Use the updated training dataset to perform semi-supervised training on the pre-trained model to obtain the updated segmentation model; S7. Using the updated segmentation model as the new pre-trained model, repeat steps S3-S6 until iterative training is completed.

[0008] Secondly, the present invention also provides an active learning training system for a three-dimensional point cloud segmentation model, comprising: The model and training set creation module is used to provide a basic segmentation model, a basic probability model, and a training dataset. The training dataset includes labeled data and unlabeled data, as well as the real labels corresponding to the labeled data. The pre-training module is used to perform supervised pre-training on the basic segmentation model using the labeled data to obtain a pre-trained model. The joint probability modeling module is used to perform forward inference on the training dataset using the pre-trained model, extract embedded features, and input the embedded features into the basic probability model for fitting training to obtain the joint probability model. The feature fusion module is used to calculate the prototype features of each segmentation category based on the output features of the pre-trained model on the labeled data; interpolate and fuse the output features of the unlabeled data with the prototype features to obtain fused features; input the fused features into the joint probability model to generate a joint probability representation; and calculate the stability score of the unlabeled data based on the volatility of the joint probability representation. A redundant point filtering module is used to remove redundancy and perform stability filtering on the unlabeled data to obtain data to be labeled, wherein the stability filtering selects a portion of the unlabeled data with lower stability scores; it receives labeling operations on the data to be labeled to form new labeled data and updates the training dataset; The model update module is used to perform semi-supervised training on the pre-trained model using the updated training dataset to obtain an updated segmentation model. The iterative training module is used to complete iterative training by using the updated segmentation model as the new pre-trained model.

[0009] Thirdly, the present invention also provides a readable storage medium storing a computer program, which, when run, executes the steps in the above-described active learning training method.

[0010] Based on the above technical solution, compared with the prior art, the beneficial effects of the present invention include at least the following: The active learning training method for the 3D point cloud segmentation model provided by this invention constructs a joint probability modeling mechanism, fits the embedded feature distribution of the basic segmentation model, and generates a joint probability distribution of unlabeled samples. By combining feature mixing and multiple inference strategies, the predictive stability of unlabeled point cloud samples is evaluated, and high-value samples are selected for labeling accordingly. This method can efficiently improve the model's discrimination ability and generalization performance under a limited labeling budget, achieving a 3D point cloud segmentation effect close to that of a fully supervised method.

[0011] The above description is merely an overview of the technical solution of the present invention. In order to enable those skilled in the art to better understand the technical means of this application and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described below in conjunction with detailed drawings. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the active learning training method provided in a typical embodiment of the present invention. Figure 2 This is a schematic diagram of the overall framework of the active learning training system provided in a typical embodiment of the present invention; Figure 3 This is a schematic diagram of the modular architecture of an active learning training system provided in a typical embodiment of the present invention. Detailed Implementation

[0013] In view of the shortcomings of the prior art, the inventors of this invention, through long-term research and extensive practice, have proposed the technical solution of this invention. The following will further explain and illustrate this technical solution, its implementation process, and its principles.

[0014] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0015] Moreover, relational terms such as “first” and “second” are used merely to distinguish one component or method step from another that has the same name, without necessarily requiring or implying any such actual relationship or order between these components or method steps.

[0016] This invention provides an active learning training method for a 3D point cloud segmentation model, comprising the following steps: S1. Provide a basic segmentation model, a basic probability model, and a training dataset, wherein the training dataset includes labeled data and unlabeled data, and the true labels corresponding to the labeled data; Specifically, the basic segmentation model is used for the final inference test of the point cloud data, the basic probability model is used to model the feature distribution of the point cloud data, and the ratio of labeled to unlabeled data in the training is preset by the task.

[0017] S2. Supervised pre-training of the basic segmentation model is performed using the labeled data to obtain a pre-trained model; Specifically, after performing strong enhancement operations on the labeled data, the data is input into the basic segmentation model, and the model parameters are optimized and trained by calculating supervised loss until the loss converges.

[0018] S3. Use the pre-trained model to perform forward inference on the training dataset, extract the embedded features, and input the embedded features into the basic probability model for fitting training. For example, the model parameters can be updated by minimizing the fitting loss of the basic probability model and using gradient descent to obtain the joint probability model. S4. Calculate the prototype features of each segmentation category based on the output features of the pre-trained model on the labeled data; interpolate and fuse the output features of the unlabeled data with the prototype features to obtain fused features; input the fused features into the joint probability model to generate a joint probability representation; calculate the stability score of the unlabeled data based on the volatility of the joint probability representation, wherein the higher the volatility, the lower the stability score, and the two are negatively correlated. S5. Redundancy removal and stability screening are performed on the unlabeled data to obtain the data to be labeled, wherein the stability screening selects a portion of the unlabeled data with lower stability scores; the labeling operation on the data to be labeled is received to form new labeled data, and the training dataset is updated; S6. Use the updated training dataset to perform semi-supervised training on the pre-trained model to obtain the updated segmentation model; S7. Using the updated segmentation model as the new pre-trained model, repeat steps S3-S6 until iterative training is completed.

[0019] In some implementations, step S2 specifically includes the following process: Strong augmentation processing is performed on the labeled data, and the enhanced labeled data is input into the basic segmentation model to obtain the corresponding segmentation prediction result; based on the segmentation prediction result and the corresponding real label, supervised loss is calculated, and the parameters of the basic segmentation model are updated according to the supervised loss; The formula for calculating the supervised loss is as follows: ; in, This indicates the supervised loss mentioned above; The representative indicates the labeled data. A set; The real label refers to the labeled data. The corresponding one-hot encoded vector; The segmentation prediction result represents the basic segmentation model's application to labeled data. The predicted class probability distribution is output.

[0020] In some implementations, the underlying probability model includes multiple coupling layers and at least one linear layer, wherein the coupling layers estimate the prior probability distribution and the linear layer estimates the posterior probability, specifically as follows: ; ; in, This represents the embedded feature; This refers to the coupling layer; Indicates the linear layer; This represents the intermediate variable output by the coupling layer; This represents the prediction of the linear layer; This represents a prior probability estimate of the embedded features of the input; This represents the posterior probability estimate of the embedded features of the input.

[0021] In some implementations, the fitting loss for training the basic probability model includes the maximum likelihood loss constructed by the coupling layer and the classification loss output by the linear layer, defined as follows: ; in, represents the total fitting loss of the basic probability model; represents the negative log-likelihood loss, which models the distribution of embedded features through the coupling layer to achieve an accurate estimate of the feature generation probability; This represents the cross-entropy classification loss, used to improve the model's discriminative ability; This represents a weighting factor used to balance the generation and modeling of input features with the discriminative task.

[0022] In some implementations, the prototype feature is calculated as follows: ; in, This indicates that the basic segmentation model applies to labeled data. The predicted label, which is the true label; This indicates that the basic segmentation model applies to labeled data. Embedded features; Indicates the category number; Indicate category Prototype features; The interpolation fusion process of the fusion feature is expressed as follows: ; in, Indicates data from unlabeled data The resulting feature vector; Indicates the total number of categories; Indicates the interpolation factor; This indicates the fusion feature.

[0023] In some implementation schemes, step S5 specifically includes the following process: S5-1: Select the top K% with the lowest stability scores from the unlabeled data as the initial candidate set through Top-K sampling; S5-2: A distance-based sampling strategy is adopted, and the K-nearest neighbor algorithm is used to remove redundant points that are spatially close and have high scores to obtain a spatially uniform subset. S5-3: Perform similarity-based sampling. By calculating the cosine similarity with the prototype features, remove data with highly similar features and high confidence in the spatially uniform subset to obtain the data to be labeled. This step is based on similarity sampling. By calculating the cosine similarity with the category prototype, remove points with highly similar features and high confidence, and finally obtain a widely distributed and diverse set of points to be labeled.

[0024] In some implementations, the semi-supervised training process includes: Construct a teacher model corresponding to the pre-trained model, and input the training dataset that has undergone weak augmentation into the teacher model, and input the training dataset that has undergone strong augmentation into the pre-trained model; The feature consistency loss, cluster-level contrast loss, and sub-cluster-level contrast loss are calculated by combining the prediction results of the teacher model and the pre-trained model, and together with the supervised loss, they form the total loss function to update the parameters of the pre-trained model.

[0025] Based on the above implementation methods, the preferred embodiment of the present invention introduces a cluster-aware contrast regularization module on the basis of the above high-value unlabeled sample screening method. It uses the distribution information between the category prototype and the sample to perform feature alignment and enhancement, and further mines the structural features in the unlabeled data.

[0026] In some implementations, the teacher model and the pre-trained model have the same network structure. In each training iteration, the parameters of the teacher model are updated using the parameters of the pre-trained model, as shown below: ; in, The parameters of the pre-trained model are represented; These represent the parameters of the corresponding teacher model; This indicates an update to the scaling factor; And / or, the total loss function is expressed as: ; Among them, among them, This represents the total loss value; This represents the feature consistency loss, used to constrain the feature output of the teacher model and the pre-trained model to remain consistent on augmented data; This represents cluster-level contrast loss, used to enhance the discriminative power between different segmentation categories; This represents the sub-cluster level contrast loss term, used to characterize the local structural differences within any segmentation category; This represents the weighting hyperparameter that balances the contributions of each loss term.

[0027] Furthermore, in some implementations, the iterative process in the training method can be performed in batches, that is, the training dataset is divided into multiple subsets, and each subset is used as a training batch to participate in model updates sequentially. The number of data samples in each batch can be any positive integer greater than 1, such as 2 to 128 samples, or it can be 1 sample, which is equivalent to training sample by sample.

[0028] Lieutenant General Figure 1 and Figure 2 As shown, as a typical application example of the above technical solution, the process of the above active learning training method can be as follows: (1) First, in joint probabilistic modeling, the loss function consists of two parts, which are applied to the coupling layer and the linear layer, respectively. Among them, the maximum likelihood estimation loss... Used to train the coupled layer. Input embedded features Mapped to latent variables through a series of reversible transformations of coupling layers. Simultaneously, the Jacobian determinant of the transformation is calculated, and then, based on the Gaussian density function, the following is obtained: log probability density The maximum likelihood estimation loss is defined as its negative log-expected value: ; Similarly, for the classification loss of a linear layer, it directly affects the latent variables output by the coupled layer. Output category prediction distribution And distribute this to segmentation prediction labels Alignment, then cross-entropy function Calculate the loss: ; (2) In the feature mixing stage, for each unlabeled sample's embedded features Each feature is interpolated and mixed with the prototype features of each category to construct a hybrid feature representation, which is then input into the basic probability model. Repeated inference. Each inference outputs the corresponding prior probability and posterior probability, denoted as [missing information]. and This aggregates the posterior predicted probabilities corresponding to all categories, i.e., averages all the mixed prediction distributions: ; ; ; in, This represents the joint probability distribution obtained through feature mixing and probabilistic inference. Based on this distribution, the prediction stability score for unlabeled samples is defined as follows: ; in, and The corresponding highest and second highest predicted probabilities in the joint probability distribution, This indicates the unlabeled sample. The predicted stability score. This stability score... This score reflects the consistency of the model's predictions for a sample under multiple feature perturbations. A higher score indicates that the model's predictions for that sample are more concentrated, suggesting stronger stability; conversely, a lower score indicates that the sample is near the class boundary or has greater uncertainty.

[0029] (3) In the redundant point filtering algorithm, the first step is to select the points with the lowest stability scores from the unlabeled points. The first step involves selecting samples to form an initial candidate set. The second step is to filter based on spatial distance, searching for each candidate point using KNN. The first step involves selecting the nearest neighbors. If the distance between two points is less than a set radius, the point with the lower stability score is retained, and the other is removed, thus obtaining a spatially more evenly distributed coverage subset. The third step is to further filter based on feature similarity, calculating the stability score of each candidate point in the coverage subset. eigenvectors With all category prototypes The cosine similarity between them is in the form of: ; The pseudo-label and confidence level are determined based on the similarity distribution between each point and the prototype. If a pair of points has the same pseudo-label and the confidence level is higher than the preset threshold, the one with the lower stability score is retained.

[0030] (4) After probabilistic modeling, feature mixing and redundant point filtering, a candidate point set with spatially dispersed distribution and high feature diversity is obtained. According to the preset annotation budget for each round, unlabeled samples with low stability scores are selected for annotation (e.g., manual annotation). After annotation is completed, the labeled dataset is updated for the training of the basic segmentation model in the next stage.

[0031] (5) In the sub-cluster-level contrastive learning module, firstly, based on the prior and posterior information of the categories output by the probability model, the dispersion score of each category is calculated to adaptively determine the number of its sub-clusters. Specifically, given a category... First, its dispersion is defined as follows: ; in, This indicates an indicator function that determines whether a sample's prediction belongs to a specific category. This dispersion reflects the uncertainty and dispersion of samples within a category, and is used to infer whether multiple potential subclusters exist within it. Based on this, the dispersion score is used to calculate a scaling factor. Number of subclusters mapped ,Right now .

[0032] For each sub-cluster, calculate its central eigenvector as the prototype of that sub-cluster. Then, based on all generated sub-cluster prototypes, a sub-cluster-level contrastive loss function is constructed, which can be expressed as: ; in, This represents the embedded features extracted by the base segmentation model. Indicates sample The corresponding sub-cluster prototype, It is a temperature scaling factor used to adjust the sensitivity of the similarity distribution.

[0033] (6) In cluster-level contrastive learning, for two data samples and If the predicted label of the basic segmentation model The corresponding teacher model's predicted labels The comparison is performed; if they are the same, the sample pair is considered a positive pair; otherwise, it is a negative pair. Next, the contrastive loss function is constructed as follows: ; in, This indicates the basic segmentation model for samples. The output embedding features, The teacher model representing the basic segmentation model is used to represent the sample. The output embedding features.

[0034] (7) The loss terms obtained from each module are weighted and summed to form the total loss function. The parameters of the basic segmentation model are updated by gradient descent to complete the active learning training of the current round.

[0035] (8) Finally, repeat the steps of unlabeled sample selection, redundant point filtering, pseudo-label generation and enhancement, and feature comparison learning to enter the next round of training. Repeat this process until the labeling budget is exhausted (or other termination conditions are met) to obtain the final point cloud segmentation model.

[0036] In the above training method, steps (1) to (4) are the process of actively selecting labeled data, and steps (5) to (7) are the further feature training of the basic segmentation model.

[0037] Corresponding to the above training method, this embodiment of the invention also provides an active learning training system for a 3D point cloud segmentation model, which includes: The model and training set creation module is used to provide a basic segmentation model, a basic probability model, and a training dataset. The training dataset includes labeled data and unlabeled data, as well as the real labels corresponding to the labeled data. The pre-training module is used to perform supervised pre-training on the basic segmentation model using the labeled data to obtain a pre-trained model. The joint probability modeling module is used to perform forward inference on the training dataset using the pre-trained model, extract embedded features, and input the embedded features into the basic probability model for fitting training to obtain the joint probability model. The feature fusion module is used to calculate the prototype features of each segmentation category based on the output features of the pre-trained model on the labeled data; interpolate and fuse the output features of the unlabeled data with the prototype features to obtain fused features; input the fused features into the joint probability model to generate a joint probability representation; and calculate the stability score of the unlabeled data based on the volatility of the joint probability representation. A redundant point filtering module is used to remove redundancy and perform stability filtering on the unlabeled data to obtain data to be labeled, wherein the stability filtering selects a portion of the unlabeled data with lower stability scores; it receives labeling operations on the data to be labeled to form new labeled data and updates the training dataset; The model update module is used to perform semi-supervised training on the pre-trained model using the updated training dataset to obtain an updated segmentation model. The iterative training module is used to complete iterative training by using the updated segmentation model as the new pre-trained model.

[0038] This invention also provides a readable storage medium storing a computer program, which, when run, executes the steps of the active learning training method provided in any of the above embodiments.

[0039] The technical solution of the present invention will be further described in detail below through several embodiments and in conjunction with the accompanying drawings. However, the selected embodiments are only for illustrating the present invention and do not limit the scope of the present invention. The following embodiments further employ the training method and system described above, construct a model based on a publicly available point cloud dataset using the training method proposed in this invention, and conduct comparative experiments with existing representative active learning methods.

[0040] Example 1 In the active learning training process, the active learning framework is set to 5 iterations, with the total annotation budget evenly distributed across each iteration. In each iteration, for each scene in the training set, 1 / 5 of the labeled points are selected according to a preset annotation ratio until all annotation budgets are exhausted. The joint probabilistic model uses the AdamW optimizer and is trained on all datasets using a cosine decay strategy, with an initial learning rate of 0.01. The coupling layer is implemented based on the FrEIA library, with a default layer count of 8; the linear layer consists of two fully connected layers. The basic segmentation model uses the SGD optimizer and a cosine decay strategy, with the initial learning rate set as follows: 0.1 for the indoor S3DIS dataset and 0.025 for the outdoor SemanticKITTI dataset. Related parameter settings are as follows: weight factor... Set to 0.2, number of neighboring points Set to 128. In the feature fusion module, the number of iterations... Set to 5 on the S3DIS dataset and 3 on the SemanticKITTI dataset. Set the Top-K sampling ratio to 5%, and the expansion coefficient... Set the momentum factor to 10. Set to 0.99. The weighting coefficients for each loss term are set as follows: , , .

[0041] Specifically, the experiment compared the model training performance under different annotation ratios, as shown below: This embodiment illustrates a comparison between the training method provided by this invention and existing point cloud active learning training methods.

[0042] Table 1 shows the performance comparison results of the 3D point cloud active learning training method described in this embodiment and existing point-level active learning training methods on the indoor S3DIS dataset. The S3DIS dataset contains 271 indoor scenes, covering 13 object categories. The third column indicates the proportion of labeled points used by the corresponding method; the fourth column is the segmentation accuracy (i.e., average intersection-union score, in percentage) of each method across all categories. As can be seen from the data in the table, under the same annotation budget settings, the method used in this embodiment achieves a higher value in the segmentation accuracy index, demonstrating its effectiveness under low annotation cost conditions.

[0043] Table 1. Performance Comparison Results of Different Active Learning Training Methods on the Indoor S3DIS Dataset

[0044] Example 2 This embodiment illustrates a comparison between the training method provided by this invention and existing point cloud active learning training methods.

[0045] Table 2 shows the performance comparison results of the 3D point cloud active learning training method described in this embodiment and existing methods on the SemanticKITTI dataset. The SemanticKITTI dataset contains 22 city street scene sequences with a total of 19 semantic categories, suitable for point cloud semantic segmentation tasks in mobile scenes. In this experiment, the total annotation budget was set to 5%, which was evenly distributed across 5 rounds of active learning iterations. In each round, 1% of the total budget of point clouds in the training set were selected for annotation and the model was updated. As can be seen from the results in the table, under different annotation ratios for different rounds, the method proposed in this embodiment generally outperforms other methods in terms of average segmentation accuracy, verifying the effectiveness and performance advantages of the proposed method under low annotation costs.

[0046] Table 2. Performance Comparison Results of Different Active Learning Training Methods on the Outdoor SemanticKITTI Dataset

[0047] Based on the above embodiments, it is clear that the active learning training method provided by the embodiments of the present invention, through a sample selection strategy that combines joint probabilistic modeling and feature mixing, sequentially selects unlabeled samples with learning value for labeling. Even with an extremely limited labeling budget, it still achieves excellent segmentation performance, verifying its feasibility and promotional value in practical applications.

[0048] On the other hand, the present invention also provides a readable storage medium having a computer program stored thereon, which, when run, implements the steps in the training method of the model provided in the above embodiments.

[0049] In another aspect, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the computer program, when run by the processor, performs the steps in the training method of the above-described model.

[0050] 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-including 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.

[0051] More specific examples (a non-exhaustive list) of readable storage media include: electrical connections (electronic devices) with 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, readable storage 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.

[0052] 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.

[0053] All aspects, embodiments, features, and examples of this invention are to be regarded as illustrative in all respects and are not intended to limit the invention, the scope of which is defined only by the claims. Other embodiments, modifications, and uses will become apparent to those skilled in the art without departing from the spirit and scope of the invention as claimed.

[0054] The use of headings and sections in this invention is not intended to limit the invention; each section can be applied to any aspect, embodiment or feature of the invention.

Claims

1. An active learning training method for a 3D point cloud segmentation model, characterized in that, include: S1. Provide a basic segmentation model, a basic probability model, and a training dataset, wherein the training dataset includes labeled data and unlabeled data, and the true labels corresponding to the labeled data; S2. Supervised pre-training of the basic segmentation model is performed using the labeled data to obtain a pre-trained model; S3. Use the pre-trained model to perform forward inference on the training dataset, extract the embedding features, and input the embedding features into the basic probability model for fitting training to obtain the joint probability model; S4. Calculate the prototype features of each segmentation category based on the output features of the pre-trained model on the labeled data; interpolate and fuse the output features of the unlabeled data with the prototype features to obtain fused features; input the fused features into the joint probability model to generate a joint probability representation; calculate the stability score of the unlabeled data based on the fluctuation of the joint probability representation. S5. Redundancy removal and stability screening are performed on the unlabeled data to obtain the data to be labeled, wherein the stability screening selects a portion of the unlabeled data with lower stability scores; the labeling operation on the data to be labeled is received to form new labeled data, and the training dataset is updated; S6. Use the updated training dataset to perform semi-supervised training on the pre-trained model to obtain the updated segmentation model; S7. Using the updated segmentation model as the new pre-trained model, repeat steps S3-S6 until iterative training is completed.

2. The active learning training method according to claim 1, characterized in that, Step S2 specifically includes: Strong augmentation processing is performed on the labeled data, and the enhanced labeled data is input into the basic segmentation model to obtain the corresponding segmentation prediction result; based on the segmentation prediction result and the corresponding real label, supervised loss is calculated, and the parameters of the basic segmentation model are updated according to the supervised loss; The formula for calculating the supervised loss is as follows: ; in, This indicates the supervised loss mentioned above; The representative indicates the labeled data. A set; The real label refers to the labeled data. The corresponding one-hot encoded vector; The segmentation prediction result represents the basic segmentation model's application to labeled data. The output category probability distribution for prediction.

3. The active learning training method according to claim 1, characterized in that, The basic probability model includes multiple coupling layers and at least one linear layer, wherein the coupling layers estimate the prior probability distribution, and the linear layer estimates the posterior probability, specifically as follows: ； ； in, This represents the embedded feature; This refers to the coupling layer; Indicates the linear layer; This represents the intermediate variable output by the coupling layer; This represents the prediction of the linear layer; This represents a prior probability estimate of the embedded features of the input; This represents the posterior probability estimate of the embedded features of the input.

4. The active learning training method according to claim 3, characterized in that, The fitting loss for training the basic probability model includes the maximum likelihood loss constructed by the coupling layer and the classification loss output by the linear layer, and is defined as follows: ； in, This represents the total fitting loss of the basic probability model; This represents the negative log-likelihood loss, which models the distribution of embedded features through a coupling layer, thereby achieving an accurate estimate of the feature generation probability. This represents the cross-entropy classification loss, used to improve the model's discriminative ability; This represents the weighting factor.

5. The active learning training method according to claim 1, characterized in that, The calculation method for the prototype feature is expressed as follows: ； in, This indicates that the basic segmentation model applies to labeled data. The predicted label, which is the true label; This indicates that the basic segmentation model applies to labeled data. Embedded features; Indicates the category number; Indicates category Prototype features; The interpolation fusion process of the fusion feature is expressed as follows: ； in, Indicates data from unlabeled data The resulting feature vector; Indicates the total number of categories; Indicates the interpolation factor; This indicates the fusion feature.

6. The active learning training method according to claim 1, characterized in that, Step S5 specifically includes: The top K percent of the unlabeled data with the lowest stability scores were selected as the initial candidate set through Top-K sampling. A distance-based sampling strategy is adopted, and the K-nearest neighbor algorithm is used to remove redundant points that are spatially close and have high scores, so as to obtain a spatially uniform subset. Similarity-based sampling is performed. By calculating the cosine similarity with the prototype features, data with high confidence and highly similar features in the spatially uniform subset are removed to obtain the data to be labeled.

7. The active learning training method according to claim 1, characterized in that, The semi-supervised training process includes: Construct a teacher model corresponding to the pre-trained model, and input the training dataset that has undergone weak augmentation into the teacher model, and input the training dataset that has undergone strong augmentation into the pre-trained model; The feature consistency loss, cluster-level contrast loss, and sub-cluster-level contrast loss are calculated by combining the prediction results of the teacher model and the pre-trained model, and together with the supervised loss, they form the total loss function to update the parameters of the pre-trained model.

8. The active learning training method according to claim 7, characterized in that, The teacher model has the same network structure as the pre-trained model. In each training iteration, the parameters of the teacher model are updated using the parameters of the pre-trained model, as shown below: ； in, The parameters of the pre-trained model are represented; These represent the parameters of the corresponding teacher model; This indicates an update to the scaling factor; And / or, the total loss function is expressed as: ； Among them, among them, This represents the total loss value; This represents the feature consistency loss, used to constrain the feature output of the teacher model and the pre-trained model to remain consistent on augmented data; This represents cluster-level contrast loss, used to enhance the discriminative power between different segmentation categories; This represents the sub-cluster level contrast loss term, used to characterize the local structural differences within any segmentation category; This represents the weighting hyperparameter that balances the contributions of each loss term.

9. An active learning training system for a 3D point cloud segmentation model, characterized in that, include: The model and training set creation module is used to provide a basic segmentation model, a basic probability model, and a training dataset. The training dataset includes labeled data and unlabeled data, as well as the real labels corresponding to the labeled data. The pre-training module is used to perform supervised pre-training on the basic segmentation model using the labeled data to obtain a pre-trained model. The joint probability modeling module is used to perform forward inference on the training dataset using the pre-trained model, extract embedded features, and input the embedded features into the basic probability model for fitting training to obtain the joint probability model. The feature fusion module is used to calculate the prototype features of each segmentation category based on the output features of the pre-trained model on the labeled data; interpolate and fuse the output features of the unlabeled data with the prototype features to obtain fused features; input the fused features into the joint probability model to generate a joint probability representation; and calculate the stability score of the unlabeled data based on the volatility of the joint probability representation. A redundant point filtering module is used to remove redundancy and perform stability filtering on the unlabeled data to obtain data to be labeled, wherein the stability filtering selects a portion of the unlabeled data with lower stability scores; it receives labeling operations on the data to be labeled to form new labeled data and updates the training dataset; The model update module is used to perform semi-supervised training on the pre-trained model using the updated training dataset to obtain an updated segmentation model. The iterative training module is used to perform iterative training by using the updated segmentation model as the new pre-trained model.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which, when run, executes the steps of the active learning training method according to any one of claims 1-8.

Images