- Original Pdf: pdf
- Keywords: point cloud, 3D computer vision, visualization
- Abstract: Recently, various networks that operate directly on point clouds have been proposed. It is of interest to us what features are utilized in those classifiers for their predictions. In this paper, we propose a novel approach to visualize important features used in classification decisions from point cloud networks. Following ideas in visualizing 2-D convolutional networks, our approach is based on gradually smoothing parts of the point cloud to remove certain shape features, and then evaluating the resulting point cloud on the original network to see whether the performance has dropped or remained the same. From these it can be seen whether certain parts are important to the point cloud classification. A main technical contribution of the paper is to propose an algorithm for smoothing point cloud shapes based on moving least squares and curvature flow. This algorithm can smoothly transition from the original point cloud to a either a uniform sphere, or a disk if the original shape is on a plane. With this algorithm, we can obtain a saliency map by adapting the Integrated-Gradients Optimized Saliency (I-GOS) algorithm, a state-of-the-art perturbation-based visualization techniques, to 3-D shapes. Experiment results revealed insights into these classifiers.