PointCNNVis: A visual analysis system for the interpretability of PointCNN.

The proposal of PointCNN enables CNN to be applied to point cloud segmentation and classification tasks. While using CNN, the local spatial information of the point cloud is also considered, which greatly improves the disorder of the point cloud. However, for many researchers, it is not clear how CNNs are applied to point clouds and how PointCNN operates. Aiming at the interpretability of PointCNN, a visual analysis method combining data flow and multi-dimensional analysis is proposed. 3D scatter plots and heatmaps are used to present the uploaded point cloud, the feature aggregation points of each layer of the neural network, and the features of each layer. Use line charts and pie charts to present the changing trend of the accuracy rate and the probability ratio of each interval during the training process, and support users to explore the flow of k-nearest neighbor points extracted from the representative point features and data of each layer. The human point cloud of 8192 raw sampling points is used as the input for the analysis. The experimental results demonstrate that the visual interaction system in this study can intuitively and efficiently explore the operation mechanism of the PointCNN network architecture. Use the PointCNN architecture to effectively explore possible problems in network segmentation tasks, and quickly find the body parts where abnormal segmentation point clouds are located. The system tool integrates a convenient interactive way to support users to explore the process of PointCNN feature extraction and the correlation between features and point clouds. At the same time, the detailed presentation of the segmentation results can effectively explore the operation mechanism and existing problems of PointCNN. The effectiveness and utility of the system tools in this paper are further validated by experimental results on collected real human point cloud data and feedback from PointCNN authors and researchers who need to use PointCNN.