Hierarchical Attention Learning of Scene Flow in 3D Point Clouds

Scene flow represents the 3D motion of every point in dynamic environments. Like optical that pixels 2D images, representation scene benefits many applications, such as autonomous driving and service robot. This paper studies problem estimation from two consecutive clouds. In this paper, a novel hierarchical neural network with double attention is proposed for learning correlation features adjacent frames refining coarse to fine layer by layer. The has new more-for-less architecture. means number input points greater than output estimation, which brings more information balances precision resource consumption. architecture, different levels generated supervised respectively. A attentive embedding module introduced aggregate using method patch-to-patch manner. proper layers supervision are carefully considered our designment. Experiments show outperforms state-of-the-art performance on FlyingThings3D KITTI Flow 2015 datasets. We also apply realistic LiDAR odometry task, an key driving. experiment results demonstrate can outperform ICP-based shows good practical application ability.