Reproduce “ICP-Flow: LiDAR Scene Flow Estimation with ICP” on OpenResearch

About this paper

Yancong Lin, Holger Caesar

Researchers at Delft University of Technology developed ICP-Flow, a learning-free method for 3D LiDAR scene flow estimation that leverages a multi-body rigid-motion assumption and a novel histogram-based ICP initialization. This approach achieves superior or competitive performance against state-of-the-art supervised and unsupervised methods on autonomous driving benchmarks, and its generated pseudo-labels enable real-time inference via a lightweight neural network.

Key insights

• Explicitly integrating a multi-body rigid-motion assumption and leveraging a histogram-based initialization within an ICP framework significantly enhances scene flow accuracy and physical consistency for dynamic objects.
• High-quality pseudo-labels derived from a principled, learning-free method can effectively enable the training of efficient, real-time neural networks for scene flow without direct reliance on human annotations.
• The proposed approach maintains robust performance even with longer temporal gaps between LiDAR scans, addressing a limitation of many existing scene flow methods.

Summary by alphaXiv.