Do Point Cloud Models Learn Object Contour Features?

The explainability and transparency of point cloud models is still an understudied topic nowadays. Although enthusiasm for designing complex and high-performance models is ongoing, little work has been devoted to decomposing and dissecting existing popular models, even those with the simplest architectures. To fill this gap, this work presents a novel observation: the point cloud models with PointNet-like structure effectively learn the contours of 3D objects in their intermediate layers. This observation presents a counterfactual for the existing view that point cloud models make decisions based on critical points. Besides, it analyzes the learned contour features of point cloud models and provides a novel inspiration for point cloud synthesis. Based on this observation, we further propose an application called “Latent Activation Maximization” (Latent AM). Latent AM is a point cloud global explainability method which requires only the classifier to be explained and generates much higher quality explanations than other non-generative model-based AMs. We believe this observation will be an important contribution to future research in the areas of point cloud network transparency as well as 3D instance synthesis.