DeFlow: Decoder of Scene Flow Network in Autonomous Driving

• DOI: 10.48550/arxiv.2401.16122
• Publication Date: 2024-01-29
• Authors: Qingwen Zhang; Yi Yang; Heng Fang; Ruoyu Geng; Patric Jensfelt

Scene flow estimation determines a scene's 3D motion field, by predicting the of points in scene, especially for aiding tasks autonomous driving. Many networks with large-scale point clouds as input use voxelization to create pseudo-image real-time running. However, process often results loss point-specific features. This gives rise challenge recovering those features scene tasks. Our paper introduces DeFlow which enables transition from voxel-based using Gated Recurrent Unit (GRU) refinement. To further enhance performance, we formulate novel function that accounts data imbalance between static and dynamic points. Evaluations on Argoverse 2 task reveal achieves state-of-the-art cloud data, demonstrating our network has better performance efficiency compared others. The code is open-sourced at https://github.com/KTH-RPL/deflow.