A cross-domain three-dimensional point cloud semantic segmentation method and system based on an active learning strategy

Abstract

The application relates to an image processing technology, and relates to a cross-domain three-dimensional point cloud semantic segmentation method and system based on an active learning strategy. The method comprises the following steps: establishing and initializing a semantic segmentation model, including a feature extractor and a classifier; training the semantic segmentation model, and optimizing parameters of the semantic segmentation model; dividing regions of point cloud data of a source domain and a target domain, calculating comprehensive scores of each region, and labeling regions with higher scores; adopting a similarity penalty mechanism to label point cloud data of a target domain region, selecting diversified regions for labeling, updating a data set, and retraining the semantic segmentation model; iteratively updating the semantic segmentation model, and processing three-dimensional point cloud data by using a final semantic segmentation model. The application reduces the demand for labeled data of the target domain, and improves segmentation performance of the model in the target domain.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to a cross-domain three-dimensional point cloud semantic segmentation method and system based on an active learning strategy. BACKGROUND

[0002] Point cloud semantic segmentation is an important part of analyzing laser radar three-dimensional point cloud data. In the field of point cloud semantic segmentation, various methods have emerged.

[0003] Before the application of machine learning, clustering algorithms such as K-means, DBSCAN and hierarchical clustering algorithm are generally used for point cloud semantic segmentation. In the existing point cloud semantic segmentation method, the clustering algorithm is often used as a segmentation tool for point cloud regions, and is an excellent tool for pre-processing point cloud data in cooperation with machine learning. However, in the processing of point cloud data lacking prior knowledge and irregular shape, the clustering algorithm has many defects.

[0004] Machine learning algorithms applied to planar or two-dimensional data processing are difficult to be directly applied to point cloud semantic segmentation. Point cloud data itself is highly discrete, and does not have the characteristics of known nearest neighbor elements like pictures; this makes the point cloud data of the same region at different angles have not small difference in shape. The model trained by using a simple neural network has poor generalization ability and cannot make the model perform reliable semantic segmentation at different angles. This problem has become a major difficulty in hindering the large-scale application of machine learning to three-dimensional point cloud semantic segmentation.

[0005] The Pointnet and Pointnet++ algorithms appeared around 2017 successfully solved this problem through structural optimization. Pointnet combines each point vector of the point cloud with other point vectors to extract global and local position information. At the same time, the K-nearest neighbor method derived from the graph convolutional neural network provides a relatively simple and reliable solution for point cloud semantic segmentation. However, this point cloud segmentation method has a large amount of calculation, and the parameter learning method is relatively rigid.

[0006] In 2019, the Transformer emerged as a unique way to perform semantic segmentation. By applying the attention mechanism of Transformer, point cloud semantic segmentation can be more efficient and accurate, and more position information and point cloud structure information can be collected at a point. The method using Transformer is superior to the traditional CNN-based method in terms of calculation amount and segmentation accuracy.

[0007] The deep learning methods mentioned above can achieve acceptable accuracy in one domain with sufficient data. However, the application scenarios of laser point clouds usually involve data from multiple domains, especially models applied to autonomous driving. If multiple domain manually annotated datasets are used, a large amount of expensive manual annotation is required. Selecting to improve the domain adaptation characteristics of the model is obviously a more cost-effective method. The domain adaptation characteristics of ordinary machine learning methods are not good, so the point cloud semantic segmentation machine learning method needs to be optimized in domain adaptation.

[0008] Therefore, methods for domain adaptation optimization have emerged. DANN is a method for improving the domain adaptation characteristics of the model in the form of a generative adversarial network. By adding a domain classification indicator to the model, the difference between the two domains in the model is narrowed. By reducing the domain classification error, the model can have good domain adaptation properties. Another method combines multiple modal data for comprehensive analysis to improve the domain adaptation ability of the model. Xmuda chooses to integrate three-dimensional point clouds and two-dimensional image data, significantly improving the generalization ability of the model.

[0009] In addition to three-dimensional point clouds, in the case of insufficient data sets or insufficient annotation resources, active learning is a method that can reduce the demand for data volume in model training. Active learning selects a part of data that is most efficient for improving the accuracy of the model after processing by using a specific strategy. There are many strategies for active learning, from the simplest random selection to the ReDAL method that integrates color discontinuity, structural complexity and softmax entropy. Selecting an excellent active learning strategy can significantly reduce the training cost of the model. SUMMARY

[0010] To solve the technical problems existing in the prior art, the present application provides a cross-domain three-dimensional point cloud semantic segmentation method and system based on an active learning strategy, which introduces a region-based segmentation strategy to select a small number of regions with the most information in the target domain for annotation, significantly reducing the demand for target domain annotation data and improving the segmentation performance of the model in the target domain.

[0011] In the embodiment of the present application, the technical scheme of the method is as follows:

[0012] S1, a semantic segmentation model is established, the established semantic segmentation model includes a feature extractor and a classifier; the feature extractor is used to extract descriptive features from the point cloud data to generate a feature representation of the point cloud data; the classifier is used to map the feature representation of the point cloud data to the class probability distribution of the point cloud data; the feature extractor and the classifier are initialized;

[0013] S2, training the semantic segmentation model, optimizing the parameters of the semantic segmentation model;

[0014] S3, region division is performed on the point cloud data of the source domain and the target domain, and the comprehensive score of each region is calculated, and the region with a higher labeling score is marked;

[0015] S4, the point cloud data of the target domain region is labeled by using a similarity penalty mechanism, and a diversified region is selected for labeling; the labeled data is integrated into the source domain data, the data set is updated, and the semantic segmentation model is retrained;

[0016] S5, iteratively updating the semantic segmentation model to obtain a final semantic segmentation model; and processing three-dimensional point cloud data by using the final semantic segmentation model.

[0017] In the embodiments of the present application, the system is realized based on the above cross-domain three-dimensional point cloud semantic segmentation method, and comprises:

[0018] A data acquisition module is configured to acquire point cloud data.

[0019] A semantic segmentation model is configured to perform semantic segmentation on the acquired point cloud data.

[0020] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0021] 1. Traditional 3D point cloud semantic segmentation methods usually rely on a large amount of labeled data, and perform poorly in cases where the source domain and the target domain have large feature differences. The present application first applies an active learning strategy to the cross-domain 3D point cloud semantic segmentation task. By selecting a small number of regions with the most information in the target domain for labeling, the demand for labeled data in the target domain is significantly reduced, and the segmentation performance of the model in the target domain is improved. This innovation solves the problem of performance degradation caused by the lack of labeled data in the current cross-domain 3D point cloud task.

[0022] 2. Traditional point cloud segmentation methods usually label and select based on the global scene. The present application introduces a region-based segmentation strategy, which divides the point cloud scene into multiple regions and takes these regions as the smallest labeling unit. Each region is independently evaluated to determine its potential contribution to the improvement of the model performance. Through this regional processing, not only the accuracy of labeling is improved, but also unnecessary data redundancy is reduced, thereby significantly reducing the labeling cost.

[0023] 3、In the annotation selection process, the present application proposes a similarity penalty mechanism, aiming to avoid selecting multiple regions with similar features. This mechanism detects regions with similar features through clustering algorithms and applies a penalty in the scoring of these regions, thus prioritizing more diverse regions for annotation. This similarity penalty strategy ensures that the selected data has a higher information content, avoiding data redundancy and annotation waste.

[0024] 4、To evaluate the annotation value of each region, the present application adopts a comprehensive multi-dimensional scoring mechanism. This scoring mechanism combines four key indicators: Softmax entropy, Maximum Mean Discrepancy (MMD), color discontinuity, and structural complexity. These indicators evaluate the uncertainty, inter-domain difference, and information content of the region from different angles, and through comprehensive scoring, they can more accurately select the regions that are most valuable for model improvement. The application of this multi-dimensional scoring mechanism in cross-domain 3D point cloud tasks significantly improves annotation efficiency and model accuracy.

[0025] 5、The present application significantly improves the cross-domain segmentation performance of the model by annotating only a small amount of data in the target domain. In experiments, using less than 10% of the target domain annotation data, the segmentation accuracy is significantly higher than traditional methods. This result shows that the present application can achieve efficient cross-domain semantic segmentation with extremely low annotation cost, suitable for various complex cross-domain scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is a flowchart of the cross-domain three-dimensional point cloud semantic segmentation method in the embodiment of the present application;

[0027] Figure 2 is a schematic diagram of the comprehensive scoring and annotation of target domain data in the embodiment of the present application. DETAILED DESCRIPTION

[0028] The present application proposes a cross-domain three-dimensional point cloud semantic segmentation method and system based on an active learning strategy, aiming to improve model performance and reduce annotation cost by intelligently selecting key data regions in the target domain. The semantic segmentation method mainly includes preliminary model training, region feature extraction, region scoring mechanism, clustering analysis and similarity penalty mechanism, and annotation and model iteration steps.

[0029] The present application will be further described in detail below in conjunction with the embodiments and drawings, but the implementation of the present application is not limited thereto.

[0030] EMBODIMENT

[0031] As Figure 1As shown, the embodiment provides a cross-domain three-dimensional point cloud semantic segmentation method based on an active learning strategy, including the following steps:

[0032] S1, a semantic segmentation model is established, and the established semantic segmentation model includes a feature extractor and a classifier; the feature extractor is used to extract descriptive features from point cloud data to generate feature representation of the point cloud data; the classifier is used to map the feature representation of the point cloud data to the class probability distribution of the point cloud data; and the feature extractor and the classifier are initialized.

[0033] The embodiment adopts an open source model to perform semantic segmentation on point cloud data, and the model mainly includes a feature extractor and a classifier. The classifier is a linear layer.

[0034] In the embodiment, the feature extractor extracts descriptive features from the point cloud data to provide input for the subsequent classifier. The feature extractor can be abstracted as a function Where x is the input point cloud data, including the geometric coordinates (x, y, z) of each point cloud data, is a parameter for controlling the feature extraction process. The function The point cloud data x is mapped to a high-dimensional feature space to generate the feature representation to describe the geometric structure and global context semantic information of the point cloud data.

[0035] To achieve this goal, the embodiment adopts the open source architecture SPVCNN as a specific implementation technique of the feature extractor. SPVCNN is a high-efficiency neural network model for point cloud data, which can efficiently extract local features and global features of point cloud data through sparse convolution. Its core mechanism includes: first, dividing the sparse point cloud data into voxel grids to capture the local geometric relationship of the point cloud data; then, extracting multi-scale features on different resolution voxel grids through sparse convolution operation; finally, integrating these multi-scale features to generate a descriptive feature vector for each point cloud data. Mathematically, SPVCNN can be described as follows:

[0036]

[0037] Where SPVCNN represents the operation logic of the sparse convolution network; is a parameter for controlling the feature extraction process. The specific parameters can include convolution kernel weight, bias, and sparse matrix configuration parameters; d represents the dimension of the feature vector.

[0038] In the initialization stage, the parameters of the feature extractor The convolution kernel weights are usually initialized as random values following a standard normal distribution to ensure that there is no initial bias in feature extraction in each direction; the bias parameters are initialized to zero to avoid artificial bias in the feature generation process. At the same time, the sparse matrix configuration of the SPVCNN is dynamically generated according to the initial distribution of the point cloud data, and is constructed by calculating the proximity relationship between points and voxels.

[0039] The classifier is used to classify the point cloud data according to the feature representation output by the feature extractor The classifier can be abstracted as a function wherein is the parameter of the classifier; the input of the classifier is the feature representation of the point cloud data obtained from the feature extractor; the output of the classifier is the probability distribution of the point cloud data belonging to each class, i.e. wherein k is the total number of classes, p i represents the probability of the point cloud data belonging to the i-th class, satisfying In the initialization stage, the parameters of the classifier are also set by a random method, with the weights initialized as small random values; and the bias is initialized to zero to ensure that the classifier has no preference for all classes in the initial state. Initially, the probability distribution output by the classifier is a uniform distribution:

[0040]

[0041] Through the initialization of the feature extractor and the classifier, the system of the embodiment can start processing the point cloud data in a neutral state and lay a foundation for subsequent parameter optimization. The feature extractor provides a high-dimensional feature representation of the point cloud data, and the classifier maps the feature representation to the class probability distribution of the point cloud data, thereby completing the preliminary point cloud classification task.

[0042] S2, training the semantic segmentation model to optimize the parameters of the semantic segmentation model.

[0043] The training process of the model is completed by using point cloud data of the source domain and the target domain , and the specific steps are as follows:

[0044] S21, input the source domain data and the target domain data into the feature extractor SPVCNN to extract the source domain feature and the target domain feature

[0045] The parameters of the feature extractor are gradually adjusted through a subsequent optimization process, so that the features extracted by it can better describe the geometric structure and global context semantic information of the point cloud data.

[0046] S22, calculate the maximum mean difference (MMD) between the source domain features and the target domain features as a feature alignment loss.

[0047] The feature alignment loss aims to minimize the difference between the source domain feature distribution and the target domain feature distribution, thereby improving the adaptability of the semantic segmentation model to the target domain point cloud data. The feature alignment loss can be calculated by the following function

[0048]

[0049] wherein is the distribution of the source domain features, is the distribution of the target domain features.

[0050] S23, input the source domain features into the classifier to predict the class probability distribution of the source domain data Compare the probability of the class to which the point cloud data belongs predicted by the classifier with the real label of the source domain Calculate the cross-entropy loss as a classification loss to measure the accuracy of the classification result:

[0051]

[0052] wherein k is the number of classes, represents the real label of the point cloud data belonging to class j (if it belongs to, the real label otherwise the real label is 0); is the probability of the point cloud data predicted by the classifier to belong to class j.

[0053] S24, integrate the feature alignment loss and the classification loss to construct the overall loss function, specifically:

[0054]

[0055] wherein λ is a weighting coefficient for balancing the relative importance of the classification loss and the feature alignment loss.

[0056] S25, update the parameters of the feature extractor and the parameters of the classifier by optimizing the overall loss function . ​

[0057] In this embodiment, the update rule of the parameters is:

[0058]

[0059] where η is the learning rate, used to control the step size of parameter update.

[0060] Through the above training steps, the feature extractor and classifier of the semantic segmentation model can gradually adapt to the characteristics of the source domain data, while improving the generalization ability for the target domain data through feature alignment.

[0061] S3, divide the point cloud data of the source domain and the target domain into regions, and calculate the comprehensive score of each region, and mark the region with a higher score.

[0062] After the preliminary training of the model is completed, the point cloud data of the target domain and the source domain is divided into regions, and the comprehensive score of the region is calculated, which includes the following steps:

[0063] S31, use the VCCS (Voxel Cloud Connectivity Segmentation) algorithm to divide the point cloud data of the source domain and the target domain, and divide the point cloud data into multiple regions and where is the nth region of the source domain, is the nth region of the target domain.

[0064] These regions will be used as the basic labeling unit, and the point cloud data in each region will be analyzed and selected as a whole, and the comprehensive score of each region will be calculated.

[0065] S32, use the trained semantic segmentation model to predict each point cloud data in the target domain, and convert the prediction output into a probability distribution through the Softmax function According to the converted probability distribution, the Softmax entropy value is calculated to measure the labeling value of the region.

[0066] The Softmax entropy value is calculated by the following formula:

[0067]

[0068] The Softmax entropy is used to measure the prediction uncertainty of the semantic segmentation model for each region, and the higher the entropy value, the higher the labeling value of the region.

[0069] S33, use the maximum mean difference (MMD) to calculate the feature difference between the regions of the source domain and the target domain, and the formula is as follows:

[0070]

[0071] wherein, for calculating the distance between the target domain region and the source domain region , which is used to measure the importance of the region to the adaptability of the cross-domain model. The calculation formula is as follows:

[0072]

[0073] wherein, is to map the target domain feature to a high-dimensional Hilbert space, is to map the source domain feature to a high-dimensional Hilbert space, and then calculate the L2 norm of the target domain feature and the source domain feature in the mapped Hilbert space. H represents the Hilbert space.

[0074] S34, calculate the color discontinuity for each target domain region .

[0075] First, the color difference between each point cloud data in the target domain region and its adjacent point cloud data is calculated. The region with larger color difference usually contains more semantic information. The specific calculation formula of the color difference is as follows:

[0076]

[0077] wherein represents the set of the k nearest points to the target domain data .

[0078] The color discontinuity of each target domain region is the average value of the color difference of each point cloud data in the region:

[0079]

[0080] S35, by calculating the surface change rate of each region in the target domain, to evaluate the structural complexity of the region.

[0081] The surface change rate is used to measure the complex boundary region and surface irregularity, and the specific calculation formula is as follows:

[0082]

[0083] wherein, represents the local surface change rate of the point .

[0084] S36, the Softmax entropy value, feature difference, color discontinuity, surface change rate are weighted and summed to obtain a comprehensive score of each region:

[0085]

[0086] Wherein, alpha, beta, gamma, delta are weight coefficients.

[0087] The comprehensive score is used to rank the importance of the regions, and the regions with higher scores are preferentially selected for labeling.

[0088] S4, the point cloud data of the target domain region is labeled using a similarity penalty mechanism to select diverse regions for labeling; the labeled data is integrated into the source domain data, the data set is updated, and the semantic segmentation model is retrained.

[0089] In the labeling selection process, the present application proposes a similarity penalty mechanism to avoid selecting multiple regions with similar features. The mechanism detects regions with similar features through a clustering algorithm and applies a penalty to the scores of these regions, thereby preferentially selecting more diverse regions for labeling.

[0090] The basic unit of labeling in this embodiment is a "region" rather than a single point cloud data, and by dividing the point cloud data into multiple regions, the workload of labeling data is significantly reduced and the labeling efficiency is improved. In order to avoid repeatedly selecting similar regions in the same labeling selection, this embodiment designs a similarity penalty mechanism, the specific steps are as follows:

[0091] S41, features are extracted for each target domain region, and a clustering algorithm (such as K-means) is used to cluster the regions, and regions with similar features are classified into the same cluster.

[0092] S42, for regions belonging to the same cluster, select according to the score ranking, but for regions ranking at the back, apply a decay factor d k-1 to reduce its score, where k represents the kth region in the cluster according to the score from high to low. In this way, the regions selected finally have diversity and similar regions are avoided.

[0093] S43, the top K regions with the highest scores selected are manually labeled, and these labeled data are integrated into the source domain data, and the semantic segmentation model is retrained using the updated data set to update the parameters of the semantic segmentation model.

[0094] S5, iteratively update the semantic segmentation model to obtain a final semantic segmentation model; use the final semantic segmentation model to process three-dimensional point cloud data.

[0095] Steps S2-S4 are repeated until a cost upper limit is reached or model performance reaches a target requirement.

[0096] The embodiment also provides a cross-domain three-dimensional point cloud semantic segmentation system based on an active learning strategy, characterized in that the system is implemented based on the cross-domain three-dimensional point cloud semantic segmentation method and comprises the following modules:

[0097] A data acquisition module is configured to acquire point cloud data.

[0098] A semantic segmentation model is configured to perform semantic segmentation on the acquired point cloud data.

[0099] Experimental results: In the S3DIS and SemanticKITTI data sets, the segmentation accuracy mIoU of the target domain point cloud under different annotation ratios is shown in Table 1 below.

[0100] Table 1

[0101]

[0102] The above embodiment is a preferred embodiment of the present application, but the embodiments of the present application are not limited to the above embodiment, and any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present application should be equivalent replacement methods and are included in the protection scope of the present application.

Claims

1. A cross-domain 3D point cloud semantic segmentation method based on an active learning strategy, characterized in that, Includes the following steps: S1. Establish a semantic segmentation model, which includes a feature extractor and a classifier. The feature extractor is used to extract descriptive features from point cloud data and generate a feature representation of the point cloud data. The classifier is used to map the feature representation of the point cloud data to the class probability distribution of the point cloud data. Initialize the feature extractor and the classifier. S2. Train the semantic segmentation model and optimize its parameters; S3. Divide the point cloud data of the source and target domains into regions, calculate the comprehensive score of each region, and mark the regions with higher scores. S4. A similarity penalty mechanism is used to label the point cloud data of the target domain region in order to select a variety of regions for labeling; The labeled data is integrated into the source domain data, the dataset is updated, and the semantic segmentation model is retrained. S5. Iteratively update the semantic segmentation model to obtain the final semantic segmentation model; use the final semantic segmentation model to process the 3D point cloud data.

2. The cross-domain 3D point cloud semantic segmentation method according to claim 1, characterized in that, Feature extractors are represented as functions Where x is the input point cloud data, containing the geometric coordinates (x, y, z) of each point cloud data. These are parameters that control the feature extraction process; they are used through functions. Point cloud data x is mapped to a high-dimensional feature space to generate feature representations. Used to describe the geometric structure and global contextual semantic information of point cloud data.

3. The cross-domain 3D point cloud semantic segmentation method according to claim 2, characterized in that, The feature extractor is implemented using the open-source architecture SPVCNN, which extracts local and global features from point cloud data through sparse convolution. First, the sparse point cloud data is divided into voxel grids to capture the local geometric relationships of the point cloud data. Then, multi-scale features are extracted on voxel grids of different resolutions through sparse convolution operations. Finally, the multi-scale features are further integrated to generate a descriptive feature vector for each point cloud data.

4. The cross-domain 3D point cloud semantic segmentation method according to claim 3, characterized in that, This includes the configuration of convolution kernel weights, biases, and sparse matrices.

5. The cross-domain 3D point cloud semantic segmentation method according to claim 1, characterized in that, The classifier is represented as a function in These are the parameters of the classifier; the input of the classifier. The feature representation of the point cloud data obtained from the feature extractor; the output of the classifier is the probability distribution of the point cloud data belonging to each category, i.e. Where k is the total number of categories.

6. The cross-domain 3D point cloud semantic segmentation method according to claim 1, characterized in that, Step S2 includes: S21. Input the source domain data and target domain data into the feature extractor respectively to extract source domain features and target domain features; S22. Calculate the maximum mean difference between the source domain features and the target domain features, and use it as the feature alignment loss; S23. Input the source domain features into the classifier to predict the category probability distribution of the source domain data; compare the probability of the category of the point cloud data predicted by the classifier with the true label of the source domain, and calculate the cross-entropy loss as the classification loss. S24. Combine feature alignment loss and classification loss to construct the overall loss function; S25. Update the parameters of the feature extractor and the classifier by optimizing the overall loss function.

7. The cross-domain 3D point cloud semantic segmentation method according to claim 1, characterized in that, Step S3 includes: S31. Use the VCCS algorithm to divide the point cloud data of the source and target domains into multiple regions. and in For the nth region of the source domain, This is the nth region of the target domain; S32. Use the trained semantic segmentation model to predict each point cloud data in the target domain, and convert the prediction output into a probability distribution using the Softmax function. Calculate the Softmax entropy value based on the transformed probability distribution; S33. Calculate the feature differences between the source and target domain regions using the maximum mean difference; S34, for each target domain region Calculate color discontinuities; S35. The structural complexity of a region is assessed by calculating the surface change rate of each region in the target domain. S36. The Softmax entropy value, feature difference, color discontinuity, and surface change rate are weighted and summed to obtain the comprehensive score for each region.

8. The cross-domain 3D point cloud semantic segmentation method according to claim 7, characterized in that, Feature differences in step S33 The calculation formula is: in, Used to calculate the target domain region and source region The distance between them; in, To incorporate target domain features Mapped to a higher-dimensional Hilbert space, To incorporate source domain features Map to a high-dimensional Hilbert space, and then calculate the L2 norm of the target domain features and source domain features in the mapped Hilbert space. H represents the Hilbert space. The calculation process for color discontinuity in step S34 is as follows: First, calculate the color difference between each point cloud data within the target domain region and its adjacent point cloud data; the color discontinuity of each target domain region is the average color difference of each point cloud data within the region.

9. The cross-domain 3D point cloud semantic segmentation method according to claim 1, characterized in that, Step S4 includes: S41. Extract features for each target domain region and use a clustering algorithm to cluster the regions, grouping regions with similar features into the same category; S42. For regions belonging to the same cluster, select them according to their score ranking, but apply a decay factor η to regions with lower rankings. z-1 Reduce its score, where z represents the z-th region of its cluster sorted from highest to lowest score; S43. Manually label the top Z regions with the highest scores, integrate the labeled data into the source domain data, retrain the semantic segmentation model using the updated dataset, and update the parameters of the semantic segmentation model.

10. A cross-domain 3D point cloud semantic segmentation system based on an active learning strategy, characterized in that, Based on the method described in any one of claims 1-9, it includes the following modules: The data acquisition module is used to acquire point cloud data; Semantic segmentation models are used to perform semantic segmentation on the acquired point cloud data.

Images