Scene flow estimation with local surface representation

Scene flow estimation aims to compute the 3D displacement field across consecutive point cloud frames and finds extensive applications in the 3D perception field. Most current methods extract features only through <i>xyz</i> coordinates, which lacks sufficient local context information, making it difficult for the network to fully understand the local spatial geometry of objects in the real scene. In addition, the importance of rotation robustness to the scene flow computational precision is ignored. Therefore, we proposed a self-supervised scene flow estimation method on the basis of surface feature encoding, which enrich local context information through Local Surface Feature Extraction (LSFE) module and made the network rotation invariant. We verify the efficiency of our approach through experimental results on KITTI and FlyingThings3D datasets. Compared to other advanced methods, our method achieves excellent results.