SplatFlow3D: Learning Scene Flow via 3D Motion Feature Splatting
Bo Wang, Dekang Zhu, Jian Li, Zhipeng Meng, Bo Wang, Dewen Hu
原文摘要
Considering the motion continuity and stability of 3D objects over short inter-frame intervals, introducing multi-frame settings in scene flow estimation can potentially address occlusion. In this paper, we propose a two-frame backbone, SimpleFlow3D, which integrates dense information from the image plane and sparse information from the depth dimension in a decoupling-followed-by-coupling manner by hierarchically constructing the 3D motion feature. SimpleFlow3D achieves SOTA results with minimal parameters. We also propose a weight-sharing multi-frame framework to splat the previous 3D motion feature to the current estimation process. The framework unifies the first and non-first frame inference by constructing pseudo inputs. After combining the proposed framework and backbone, the newly born multi-frame method SplatFlow3D further advances the state of the art. It achieves lower errors than the two-frame backbone, with fewer parameters than current methods. On both KITTI and Spring, SplatFlow3D ranked 1st over all published and submitted methods. Experiments show that SplatFlow3D improves occlusion handling, non-rigid motion estimation, and generalization to real-world scenes. Code is available at https://github.com/wwsource/SplatFlow3D.
cs.CV