Dynamic Inference Acceleration of 3D Point Cloud Deep Neural Networks Using Point Density and Entropy
This paper introduces a density- and entropy-adaptive inference acceleration method for 3D point cloud based deep neural networks. Based on the entropy of each input frame, the method first determines the number of points to be inferred. Then we apply a novel density calculation method to sample the points in the order of density of each point. Experiments on two representative 3D scene flow estimation models with the KITTI dataset show that the proposed scheme reduces inference latency by 32% each within 0.01m of the estimation error.