pGS-CAM: Interpretable LiDAR Point Cloud Semantic Segmentation via Gradient Based Localization

Abhishek Kuriyal, Vaibhav Kumar

Keywords: Deep learning, LiDAR, Point Cloud, Class activation mapping, Semantic Segmentation

TL;DR: Novel algorithm to explain and visualize point cloud semantic segmentation architectures via saliency/heat maps.

Abstract: To extract the local information required for effective semantic segmentation of point clouds, a number of deep learning architectures typically make use of sophisticated feature extractors. Unfortunately, there has not been a lot of discussion on how to interpret their forecasts, which is essential if deployed in real-world settings. To that end, we propose pGS-CAM (point cloud Grad-Seg-CAM), a quick and effective gradient-based method for class activation mapping in point cloud semantic segmentation architectures. To gain insight into what each intermediate layer of the architecture does, our technique provides a heatmap for the corresponding layer. We use the popular semantic segmentation architecture (RandLA-Net) and a commonly used MLS dataset (SemanticKITTI) for our experimentation.