TRANSPOINTFLOW: LEARNING SCENE FLOW FROM POINT CLOUDS WITH TRANSFORMER

Scene flow estimation is the task of obtaining 3D motion from a dynamic scene. Due to the sparseness of point clouds, extracting features for a local group of points separately may result in different features that may all belong to the same object. This difference makes global correlation prone to producing an unacceptable flow. Local correlation restricts the algorithm to capturing limited movements and fails when fast movement or large deformation of an object occurs. Therefore, we propose a transformer-based scene flow method that can perform global feature modeling through a self-attention layer. Moreover, we propose a cross-attention-based flow embedding layer for global feature matching. We further propose a learnable attention-based up-sampling layer to up-sample the estimated flow to higher resolution based on a single feature scale, eliminating the need to model global dependencies at all scales. Experimental results show that our model produces competitive results on Flyingthings3D and KITTI datasets with efficient performance.