Integrated Gradients for Feature Assessment in Point Cloud-Based Data Sets
Markus Schwegler, Christoph Müller, Alexander Reiterer
Abstract
Integrated gradients is an explainable AI technique that aims to explain the relationship between a model’s predictions in terms of its features. Adapting this technique to point clouds and semantic segmentation models allows a class-wise attribution of the predictions with respect to the input features. This allows better insight into how a model reached a prediction. Furthermore, it allows a quantitative analysis of how much each feature contributes to a prediction. To obtain these attributions, a baseline with high entropy is generated and interpolated with the point cloud to be visualized. These interpolated point clouds are then run through the network and their gradients are collected. By observing the change in gradients during each iteration an attribution can be found for each input feature. These can then be projected back onto the original point cloud and compared to the predictions and input point cloud. These attributions are generated using RandLA-Net due to it being an efficient semantic segmentation model that uses comparatively few parameters, therefore keeping the number of gradients that must be stored at a reasonable level. The attribution was run on the public Semantic3D dataset and the SVGEO large-scale urban dataset.