Hierarchical progressive fusion: A novel explainability method for point cloud deep neural networks

Deep neural networks have achieved strong performance in 3D point cloud understanding, yet their decision process remains difficult to interpret. Existing explainability methods are mainly designed for images and are not well suited to irregular, sparse point clouds. In point cloud networks, Farthest Point Sampling (FPS) breaks the correspondence between deep features and original inputs, leading to spatial inconsistency. Some existing methods alter the original point cloud by removing points during explanation, which alters the input itself and undermines input consistency. We propose Hierarchical Progressive Fusion (HPF), which progressively fuses multi-level gradients and activations and projects them back to the original input space to achieve spatial consistency. Furthermore, HPF only requires a single forward and backward propagation to generate heatmaps, achieving input consistency. Moreover, HPF does not modify the point cloud network architecture as a plug-and-play tool. Experiments on ShapeNet-Part and SUN RGB-D datasets show that HPF yields faithful and effective explanations for point cloud models. Code is available at https://github.com/perfect979/HPF .