3D object recognition system based on deep hough voting and recognition method thereof

Abstract

The application provides a 3D object recognition system based on deep Hough voting and a recognition method thereof. The recognition system comprises a votenet detection module, which is used for inputting N point clouds, learning point cloud features based on a pointnet++ backbone network, and generating M votes from seed points; an RGB module connected with the seed points, which is used for object detection on a 2D image and point cloud feature extraction on a point cloud, and then 2D image voting and semantic and texture prompts are lifted to 3D seed points; a PPC module connected with the seed points, which is used for analyzing feature information carried by the seed point cloud; an OOC module, which is used for forming a cluster group after the point cloud is subjected to initial processing, Hough voting and sampling clustering; and a GSC module connected with the seed points and the cluster group, which is used for applying an MLP layer to further aggregate global information, and outputting a feature map which is then expanded and combined with an output feature map of the OOC module. Compared with the prior art, the application can directly process raw data without relying on any 2D detector.

Description

Technical Field

[0001] This invention relates to a 3D object recognition system and method based on depth Hough voting. Background Technology

[0002] Current 3D object detection methods are heavily influenced by 2D detectors. In order to take advantage of the architecture in 2D detectors, they typically convert 3D point clouds into regular meshes (i.e., voxel meshes or bird's-eye view images) or rely on detections in 2D images to propose 3D bounding boxes.

[0003] To avoid handling irregular point clouds, current 3D object detection methods heavily rely on 2D-based detectors in various aspects. For example, "Deep sliding shapes for amodal 3D object detection in rgb-d images" extends 2D detection frameworks such as Fast / Mask R-CNN to 3D. They voxelize irregular point clouds into regular 3D meshes and apply 3D CNN detectors, which fail to take advantage of the sparsity in the data and suffer from high computational costs due to expensive 3D convolutions. Alternatively, "Multi-view 3D object detection network for autonomous driving" projects points onto regular 2D bird's-eye view images and then applies 2D detectors to locate objects. However, this sacrifices geometric details that may be crucial in cluttered indoor environments. Recently, "2D-driven 3D object detection in rgb-d images" proposed a cascaded two-step pipeline that first detects objects in a front-view image and then locates objects in a frustum point cloud extruded from a 2D box; however, this is strictly dependent on 2D detectors and will completely miss objects if they are not detected in 2D.

[0004] Hough voting, initially introduced in the late 1950s, transforms the problem of detecting simple patterns in point samples into detecting peaks in the parameter space. The generalized Hough transform further extends this technique to image patches as indicators of the presence of complex objects. Examples of Hough voting include the pioneering work "Robust object detection with interleaved categorization and segmentation," which introduced implicit shape models, extracted planes from 3D point clouds, and 6D pose estimation, among others. Hough voting has also been combined with advanced learning techniques before; in "Object detection using amax-margin hough transform," the vote assigns weights indicating their importance, which are learned using a maximum margin framework. Hough forests for object detection are introduced in "Hough forests for object detection, tracking, and action recognition."

[0005] Recently, the paper "Deep learning of local RGB-D patches for 3D object detection and 6D pose estimation" demonstrated an improved voting-based 6D pose estimation by constructing a codebook using extracted deep features. The classic Hough algorithm is used to extract circular patterns from car logos and then feed them into a deep classification network. There are also works using Hough voting for 3D object detection, such as "Implicit shape models for object detection in 3D point clouds," which employs a similar pipeline to 2D detectors.

[0006] Recently, there has been a surge of interest in designing deep network architectures suitable for point clouds, demonstrating significant performance in 3D object classification, object segmentation, and scene segmentation. In the context of 3D object detection, VoxelNet learns voxel feature embeddings from points in voxels, while PointNets are used to locate objects in truncated volumetric point clouds extruded from 2D bounding boxes. However, few methods have investigated how to directly extract and detect 3D objects from raw point cloud representations. Summary of the Invention

[0007] The main objective of this invention is to propose a 3D object recognition system and method based on depth Hough voting, which can directly process raw data using a point cloud focusing 3D detection framework without relying on any 2D detector.

[0008] To achieve the above objectives, the present invention provides a 3D object recognition system based on depth Hough voting, comprising:

[0009] The VoteNet detection module is used to take N point clouds as input and learn point cloud features based on the PointNet++ backbone network and learn Hough voting from seed points to generate M votes. Each vote has 3D coordinates and a high-dimensional feature vector.

[0010] The RGB module, connected to the seed point, is used for object detection on 2D images and point cloud feature extraction on point clouds. Then, the 2D image voting, semantic and texture cues are promoted to the 3D seed point. The seed point with concatenated image and point cloud features will generate a vote for the center of the 3D object.

[0011] The PPC module, connected to the seed points, is used to analyze the feature information carried by the seed point cloud.

[0012] The OOC module, after initial processing, Hough voting, and sampling clustering to form clusters, identifies K object centers. Before proposal and classification, and after max pooling, the OOC module utilizes CGNL's self-attention module to generate new feature maps that record the affinity between all clusters, i.e., the relationships between objects.

[0013] The GSC module, connected to the seed point and cluster, applies the MLP layer to further aggregate global information, and the output is subsequently expanded and combined with the output feature map of the OOC module.

[0014] Preferably, the Hough voting is determined by looking up in a pre-computed codebook and generating votes using a deep network based on the voting module.

[0015] Preferably, the voting module is implemented through a multilayer perceptron (MLP) network, which has a fully connected layer, ReLU, and batch normalization.

[0016] Preferably, the multilayer perceptron (MLP) network is used to acquire seed features f. i It also outputs the Euclidean space offset Δx. i ∈R 3 and feature offset Δf i ∈R C So that from the seed point s i The generated vote v i =[y i g i ] has y i =x i +Δx i and g i =fi +Δf i Seed point s i =[x i ;f i ], x i ∈R 3 f i ∈R C R 3 The spatial offset is represented as a three-dimensional vector, where each component is a real number, R. C The feature offset is represented as a C-dimensional vector, where each component is a real number.

[0017] Predicted Euclidean space offset Δx i Calculated from the regression loss function:

[0018]

[0019] Among them 1[s i [on object] indicates the seed point s i Whether it is located on the surface of the object, M pos These are seed points on the object's surface. From seed position x i The ground displacement to the center of the bounding box of its object.

[0020] Preferably, in the PPC module, the relationship between point patches is regarded as a first-level context, i.e., patch-patch context, and for each point patch, a self-focus module is used to gather information from all other patches before sending it to the voting stage.

[0021] Preferably, in the OOC module, the relationship between objects is regarded as a second-level context, namely object-object context.

[0022] Preferably, in the GSC module, information about the entire scene is regarded as a third-level context, namely the global scene context, in order to utilize global scene context information to improve the feature representation of 3D bounding box proposals and object classification without explicit scene supervision.

[0023] Preferably, in the GSC module, a new branch is created using inputs from seed points and clusters. By utilizing patch vectors and cluster vectors, information from all patches and candidate objects is combined. Following the multi-scale feature fusion strategy in the context modeling of 2D detectors, the patch vectors and cluster vectors are connected to form a global feature vector.

[0024] Preferably, the multi-scale feature fusion process can be summarized as follows:

[0025] Cnew =MLP([max(C); max(P)])+C OOC Where max(C) represents max pooling of the results from the OOC module, max(P) represents max pooling of the results from the PPC module, and MLP represents MLP of both max(C) and max(P). OOC This indicates that the results of the PPC-OOC operation are added together.

[0026] To achieve the above objectives, the present invention also provides a 3D object recognition method based on depth Hough voting, applied to the aforementioned recognition system, mainly including the following steps:

[0027] Step 1: Deploy the code, including the required environment, compiling PointNet++, and preparing the dataset;

[0028] Step 2: Run train.py to train the dataset. Some options are available when training the dataset, such as batch-size, epoch, and learn-rate, which can be implemented by attaching code when running train.py.

[0029] Step 3: After training is complete, the training weights will be in the log folder. At this point, you can choose to run eval.py to verify the network performance. Alternatively, you can place the dataset in the specified format, run demo.py, and generate several prediction files, including seed point cloud files, voting point cloud files, and clustering files.

[0030] Step 4: Use the MeshLab program to view the ply file generated in Step 3 in a 3D perspective.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows: The 3D object recognition system based on depth Hough voting of the present invention directly processes the raw data, does not rely on any 2D detector, and can accurately perform recognition. Attached Figure Description

[0032] Figure 1 This is a structural block diagram of the 3D object recognition system based on depth Hough voting according to the present invention.

[0033] Figure 2(a) is a schematic diagram of the original point cloud.

[0034] Figure 2(b) is a schematic diagram of the seed point.

[0035] Figure 2(c) is a schematic diagram of the center point of the object after voting.

[0036] Figure 2(d) is a schematic diagram of the 3D frame derived from the center point of the object.

[0037] Figure 3(a) is a structural diagram of the entire scene.

[0038] Figure 3(b) is the result of applying the recognition system of the present invention to recognize Figure 3(a). Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] It should be emphasized that, in describing this invention, various formulas and constraints are distinguished by consistent reference numerals, but it is not excluded that different reference numerals may be used to identify the same formulas and / or constraints. The purpose of this arrangement is to more clearly illustrate the features of this invention.

[0041] This invention proposes a point cloud-focused 3D detection framework that can directly process raw data without relying on any 2D detector, whether for architecture or object proposal. The detection network VoteNet is based on the latest advancements in 3D deep learning models for point clouds and is inspired by the generalized Hough voting process for object detection. It utilizes PointNet++—a hierarchical deep network for point cloud learning—to alleviate the need to convert point clouds into regular structures. By directly processing the point cloud, it not only avoids information loss during quantization but also leverages the sparsity of the point cloud by computing only the sensed points.

[0042] In images, there is usually a pixel near the center of an object, but this is not typically the case in point clouds. Since depth sensors only capture the surface of an object, the center of a 3D object is likely in empty space far from any point. Therefore, point-based networks struggle to aggregate scene context near the object center, and simply increasing the receptive field does not solve the problem, as it leads to more nearby objects and clutter as the network captures a larger context. To address this, this invention proposes a voting mechanism for point cloud depth networks similar to classic Hough voting. Through voting, new points near the object center are essentially generated, and these new points can be grouped and aggregated to generate box proposals.

[0043] Compared to traditional Hough voting, which involves multiple independent modules and is difficult to jointly optimize, the VoteNet network is end-to-end optimizable. Specifically, after the input point cloud is passed through the backbone network, a set of seed points is sampled, and votes are generated based on their features, with the votes specifically targeting the object centers. Therefore, vote clusters appear near the object centers and, in turn, can be aggregated through learned modules to generate box proposals. The result is a powerful 3D object detector that is purely geometric and can be directly applied to point clouds.

[0044] RGB images offer higher resolution than depth images or LiDAR point clouds and contain rich textures unavailable in the point domain. Furthermore, images can cover the "blind spots" of active depth sensors, typically caused by reflective surfaces. On the other hand, images are limited in 3D detection tasks because they lack absolute measurements of object depth and scale, which 3D point clouds provide. Therefore, this invention proposes incorporating RGB information into VoteNet.

[0045] Furthermore, because VoteNet processes each point and object independently, without considering the relationships between different objects or between an object and its surrounding scene, this limits the detection accuracy. Therefore, we added PPC, OOC, and GSC modules.

[0046] Specifically, such as Figure 1 As shown, this invention proposes a 3D object recognition system based on depth Hough voting, comprising:

[0047] The VoteNet detection module is used to take N point clouds as input and learn point cloud features based on the PointNet++ backbone network and learn Hough voting from seed points to generate M votes. Each vote has 3D coordinates and a high-dimensional feature vector.

[0048] The RGB module, connected to the seed point, is used for object detection on 2D images and point cloud feature extraction on point clouds. Then, the 2D image voting, semantic and texture cues are promoted to the 3D seed point. The seed point with concatenated image and point cloud features will generate a vote for the center of the 3D object.

[0049] The PPC module, connected to the seed points, is used to analyze the feature information carried by the seed point cloud.

[0050] The OOC module, after initial processing, Hough voting, and sampling clustering to form clusters, identifies K object centers. Before proposal and classification, and after max pooling, the OOC module utilizes CGNL's self-attention module to generate new feature maps that record the affinity between all clusters, i.e., the relationships between objects.

[0051] The GSC module, connected to the seed point and cluster, applies the MLP layer to further aggregate global information, and the output is subsequently expanded and combined with the output feature map of the OOC module.

[0052] Combination Figures 2(a) to 2(d)As shown, for the VoteNet detection module: First, from the input N point clouds, our goal is to generate M votes, each with 3D coordinates and a high-dimensional feature vector. There are two main steps: learning point cloud features through a backbone network and learning Hough votes from seed points. The mature PointNet++ network is used here for implementation. Compared to traditional Hough voting, the Hough voting (based on offsets from local keypoints) is determined by looking up in a pre-computed codebook and using a deep network based on the voting module to generate votes; it is both efficient and more accurate because it is jointly trained with the rest of the pipeline.

[0053] Then, given a set of seed points s i =[x i ;f i ], x i ∈R 3 f i ∈R C R 3 The spatial offset is represented as a three-dimensional vector, where each component is a real number, R. C The feature offset is represented as a C-dimensional vector, where each component is a real number. The shared voting module independently generates votes from each seed. Specifically, the voting module is implemented using a multilayer perceptron (MLP) network, which has fully connected layers, ReLU, and batch normalization. The MLP network is used to obtain the seed features f. i It also outputs the Euclidean space offset Δx. i ∈R 3 and feature offset Δf i ∈R C So that from the seed point s i The generated vote v i =[y i g i ] has y i =x i +Δx i and g i =f i +Δf i .

[0054] Predicted Euclidean space offset Δx i Calculated from the regression loss function:

[0055]

[0056] Among them 1[s i [on object] indicates the seed point s i Whether it is located on the surface of the object, M posThese are seed points on the object's surface. From seed position x i The ground displacement to the center of the bounding box of its object.

[0057] For the RGB module, given an RGB-D input (depth image converted to point cloud), the module initially has two separate branches: one for 2D object detection on the image and the other for point cloud feature extraction on the point cloud (using the PointNet++ backbone); then, 2D image voting, along with semantic and texture cues, is used to elevate 3D seed points (fusion part); finally, the seed points with concatenated image and point cloud features generate votes for the center of the 3D object.

[0058] This invention generates image votes based on a set of candidate boxes from a 2D detector. An image vote, geometrically speaking, is simply a vector connecting an image pixel to the center of the bounding box of the 2D object to which that pixel belongs. Each image vote is further enhanced with semantic and texture cues from features of its source pixels, resulting in a total of F′ dimensions for each image vote. To form a set of boxes for a given RGB image, this invention applies an off-the-shelf 2D detector (e.g., R-CNN), pre-trained on the color channels of an RGB-D dataset. The 2D detector outputs M most confident bounding boxes and their corresponding classes. A vote is assigned to the center of the box for each pixel within a detected box. Pixels within multiple boxes are given multiple votes (the corresponding 3D seed point is copied to each of them), while any pixels outside the box are zero-padded.

[0059] In the PPC module, the relationships between point patches are considered as a first-level context, i.e., patch-patch context. For each point patch, the basic idea is to use a self-focusing module to gather information from all other patches before sending it to the voting stage. After the initial point cloud is processed by PointNet++, a set of seed point clouds is obtained, see... Figure 1 In the PPC module, this module analyzes the feature information carried by the seed point cloud. At this level, the context information between point patches helps to alleviate the data loss problem by collecting supplementary information from similar patches. On the other hand, it considers the interrelationship between blocks used for voting by aggregating voting information from the current point block and all other blocks.

[0060] In the OOC module, most existing object detection frameworks detect each object individually. VoteNet is no exception, where each cluster is independently fed into a multilayer perceptron (MLP) network to regress its object class and bounding box. However, combining features from other objects provides more information about object relationships, which has proven helpful in image object detection. Intuitively, objects receive weighted messages from those highly related objects, so that the final predicted object is determined not only by its own individual feature vectors but also by object relationships. Therefore, this invention treats the relationships between objects as a second-level context, namely, object-object context.

[0061] like Figure 1 As shown in the OOC module, after the point cloud undergoes initial processing, Hough voting, and sampling clustering to form clusters, there are K object centers. Before proposal and classification, and after max pooling, the OOC module uses CGNL's self-attention module to generate new feature maps to record the affinity between all clusters, i.e., the relationship between objects.

[0062] In the GSC module, the entire point cloud typically contains rich scene context information, which helps improve the accuracy of object detection. For example, when the entire scene is a restaurant rather than a bathroom, it is more likely that a chair will be identified instead of a toilet. Therefore, this invention treats information about the entire scene as a third-level context, namely the Global Scene Context (GSC), to leverage global scene context information to improve the feature representation of 3D bounding box proposals and object classification without explicit scene supervision.

[0063] In the GSC module, a new branch is created using inputs from seed points and clusters. Leveraging patch vectors and cluster vectors, information from all patches and candidates is combined, following the multi-scale feature fusion strategy in the context modeling of 2D detectors. Patch vectors and cluster vectors are concatenated to form a global feature vector. Then, a multilayer perceptron (MLP) network is applied to further aggregate the global information, and the output is subsequently expanded and combined with the feature map output from the OOC module.

[0064] The multi-scale feature fusion process can be summarized as follows:

[0065] C new =MLP([max(C); max(P)])+C OOC

[0066] Where max(C) represents max pooling of the results from the OOC module, max(P) represents max pooling of the results from the PPC module, and MLP represents MLP of both max(C) and max(P). OOCThis indicates that the results of the PPC-OOC operation are added together.

[0067] refer to Figure 1 The overall framework diagram shown illustrates that the final result is obtained by adding the results from two branches: one branch represents max pooling of the results from the OOC module, max pooling of the results from the PPC module, and MLP of the results from max(P) and max(C); the other branch is the result directly obtained after the PPC-OOC operation, i.e., C. OOC Add the two together.

[0068] As shown in Figures 3(a) and 3(b), there are always connections between objects and between objects and the scene. Therefore, it is difficult to distinguish what an object is when it appears alone. However, if global scene information is added, for example, when the scene is known to be a restaurant, the probability of guessing the detected object as a chair reaches 60%. If the correlation information between objects is added, when the surrounding objects are known to be tables and chairs, the probability of guessing the detected object as a chair reaches 85%. Combining the two, the multi-level information will make the accuracy rate reach over 90%, which is the significance of this invention. As can be seen from Figures 3(a) and 3(b), almost all tables, chairs, and bookshelves can be identified, with excellent results.

[0069] This invention also proposes a 3D object recognition method based on depth Hough voting, applied to the aforementioned recognition system, which mainly includes the following steps:

[0070] Step 1: Deploy the code, including the required environment, compiling PointNet++, and preparing the dataset;

[0071] Step 2: Run train.py to train the dataset. Some options are available when training the dataset, such as batch-size, epoch, and learn-rate, which can be implemented by attaching code when running train.py.

[0072] Step 3: After training is complete, the training weights will be in the log folder. At this point, you can choose to run eval.py to verify the network performance, or you can run demo.py after placing the dataset in the specified format and generate several prediction files, including seed point cloud files, voting point cloud files and clustering files.

[0073] Step 4: Use the MeshLab program to view the ply file generated in Step 3 in a 3D perspective.

[0074] Furthermore, this invention can also be applied to robotic vacuum cleaners. It mainly uses point cloud data acquired by the binocular camera of the robotic vacuum cleaner to identify obstacles through sampling, Hough voting, clustering, and recognition. At the same time, it uses 2D images acquired by the built-in RGB camera to assist 3D recognition. In addition, related modules such as patch-patch, object-object, and global-scene are added before Hough voting to improve the recognition accuracy.

[0075] This is because, with the continuous development of the times and the continuous improvement of technology, people's lives are getting better and better, and large-scale robotic vacuum cleaners have become essential products in shopping malls, transportation hubs, and other places. However, most robotic vacuum cleaners on the market are currently manually operated, which takes a lot of time and manpower to clean a large area. Based on this idea, an intelligent robotic vacuum cleaner was developed. The obstacle avoidance radar of this intelligent robotic vacuum cleaner is located at the bottom. Because the robot needs to include equipment such as a water tank, its own limitations result in an excessively tall body. When there is no space below the obstacle, but the obstacle is located at the top of the robot (such as a table or chair), the radar will fail to avoid the obstacle. Based on this problem, there are two solutions: one is to install another radar at the top, and the other is to apply the recognition system of this invention, which only requires a binocular camera. Currently, most of the products on the market use the first method. However, based on cost-effectiveness considerations, the recognition system of this invention can significantly reduce costs, because binocular cameras are much cheaper than radar. Therefore, the recognition system and recognition method of this invention have a simpler structure and a wider range of applications.

[0076] In summary, the 3D object recognition system based on depth Hough voting of the present invention directly processes raw data, does not rely on any 2D detector, and can accurately perform recognition.

[0077] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A 3D object recognition system based on depth Hough voting, characterized in that, include: The VoteNet detection module is used to take N point clouds as input and learn point cloud features based on the PointNet++ backbone network and learn Hough voting from seed points to generate M votes. Each vote has 3D coordinates and a high-dimensional feature vector. The RGB module, connected to the seed point, is used for object detection on 2D images and point cloud feature extraction on point clouds. Then, the 2D image voting, semantic and texture cues are promoted to the 3D seed point. The seed point with concatenated image and point cloud features will generate a vote for the center of the 3D object. The PPC module, connected to the seed points, is used to analyze the feature information carried by the seed point cloud. The OOC module, after the point cloud has undergone initial processing, Hough voting, and sampling clustering to form clusters, has K object centers. Before proposal and classification and after max pooling, the OOC module uses CGNL's self-attention module to generate new feature maps to record the affinity between all clusters, i.e. the relationship between objects. as well as The GSC module, connected to the seed point and cluster, applies the MLP layer to further aggregate global information, and then outputs an extension that is combined with the output feature map of the OOC module. In the GSC module, a new branch is created using inputs from seed points and clusters. Utilizing patch vectors and cluster vectors, information from all patches and candidate objects is combined. Following the multi-scale feature fusion strategy in the context modeling of 2D detectors, the patch vectors and cluster vectors are concatenated to form a global feature vector. This multi-scale feature fusion process can be summarized as follows: in, This indicates that the results from the OOC module are max-pooled. This indicates max pooling of the results from the PPC module, while MLP indicates max pooling of the results from the PPC module. and Both are used in MLP. This indicates that the results of the PPC-OOC operation are added together.

2. The 3D object recognition system based on depth Hough voting according to claim 1, characterized in that: The Hough voting is determined by looking up in a pre-computed codebook and generating votes using a deep network based on the voting module.

3. The 3D object recognition system based on depth Hough voting according to claim 2, characterized in that: The voting module is implemented through a multilayer perceptron (MLP) network, which has a fully connected layer, ReLU, and batch normalization.

4. The 3D object recognition system based on depth Hough voting according to claim 3, characterized in that: The multilayer perceptron (MLP) network is used to acquire seed features. It also outputs the Euclidean space offset. ∈ and feature offset ∈ This makes the seed point generated votes = [ ; [Has] = + and = + Seed point , = [ ; ], ∈ , ∈ , The spatial offset is represented as a three-dimensional vector, where each component is a real number. The feature offset is represented as a C-dimensional vector, where each component is a real number. Predicted Euclidean space offset Calculated from the regression loss function: in Indicator Seed Point Whether it is located on the surface of the object, These are seed points on the object's surface. From seed position The ground displacement to the center of the bounding box of its object.

5. The 3D object recognition system based on depth Hough voting according to claim 1, characterized in that: In the PPC module, the relationship between point patches is regarded as a first-level context, namely patch-patch context, and for each point patch, a self-focus module is used to gather information from all other patches before sending it to the voting stage.

6. The 3D object recognition system based on depth Hough voting according to claim 1, characterized in that: In the OOC module, the relationships between objects are considered as a second-level context, namely object-object context.

7. The 3D object recognition system based on depth Hough voting according to claim 1, characterized in that: In the GSC module, information about the entire scene is treated as a third-level context, namely the global scene context, to leverage global scene context information to improve feature representations for 3D bounding box proposals and object classification without explicit scene supervision.

8. A 3D object recognition method based on depth Hough voting, characterized in that: The identification system applied to any one of claims 1-7 mainly includes the following steps: Step 1: Deploy the code, including the required environment, compiling PointNet++, and preparing the dataset; Step 2: Run train.py to train the dataset. When training the dataset, the batch-size, epoch, and learn-rate options are optional and can be implemented by attaching code when running train.py. Step 3: After training is complete, the training weights will be in the log folder. At this point, you can choose to run eval.py to verify the network performance. Alternatively, you can place the dataset in the specified format, run demo.py, and generate several prediction files, including seed point cloud files, voting point cloud files, and clustering files. Step 4: Use the MeshLab program to view the ply file generated in Step 3 in a 3D perspective.

Images