No Pain, Big Gain: Classify Dynamic Point Cloud Sequences with Static Models by Fitting Feature-level Space-time Surfaces

Scene flow is a powerful tool for capturing the motion field of 3D pointclouds. However, it is difficult to directly apply flow-based models to dynamicpoint cloud classification since the unstructured points make it hard or evenimpossible to efficiently and effectively trace point-wise correspondences. Tocapture 3D motions without explicitly tracking correspondences, we propose akinematics-inspired neural network (Kinet) by generalizing the kinematicconcept of ST-surfaces to the feature space. By unrolling the normal solver ofST-surfaces in the feature space, Kinet implicitly encodes feature-leveldynamics and gains advantages from the use of mature backbones for static pointcloud processing. With only minor changes in network structures and lowcomputing overhead, it is painless to jointly train and deploy our frameworkwith a given static model. Experiments on NvGesture, SHREC'17, MSRAction-3D,and NTU-RGBD demonstrate its efficacy in performance, efficiency in both thenumber of parameters and computational complexity, as well as its versatilityto various static backbones. Noticeably, Kinet achieves the accuracy of 93.27%on MSRAction-3D with only 3.20M parameters and 10.35G FLOPS.