Reproduce “DeFlow: Decoder of Scene Flow Network in Autonomous Driving” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qingwen Zhang, Yi Yang, Heng Fang, Ruoyu Geng, Patric Jensfelt
DeFlow, a scene flow estimation network for autonomous driving, addresses real-time processing of large point clouds by introducing a GRU-based decoder to recover point-specific features from voxelized data and an improved loss function for static-dynamic point imbalance. It achieved state-of-the-art performance on the Argoverse 2 dataset with an EPE 3-Way of 0.0534 while operating at real-time speeds of over 20 FPS.
Key insights
- Iterative refinement using a Gated Recurrent Unit (GRU) in the decoder is effective in recovering point-specific feature differentiation lost during voxelization, improving overall scene flow accuracy.
- A balanced loss formulation that sums errors across distinct motion speed categories provides superior training signals compared to single scaling factors, mitigating data imbalance and enhancing dynamic point accuracy.
- Competitive scene flow accuracy can be achieved even with coarser voxel resolutions when combined with an advanced decoder, leading to significant improvements in computational efficiency and real-time performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.