Weakly Supervised Cross-Modal Learning for 4D Radar Scene Flow Estimation

Researchers proposed a weakly supervised framework for 4D radar scene flow that learns from 2D tracking, segmentation, and vehicle odometry.

The method avoids heavy reliance on LiDAR and manual annotations by using cross-modal supervision to guide radar motion estimation.

It achieved state-of-the-art performance on real-world data, including the View-of-Delft dataset.

The approach could make autonomous radar sensing more practical, scalable, and cost-effective.