Reproduce “TARS: Traffic-Aware Radar Scene Flow Estimation” on OpenResearch

About this paper

Jialong Wu, Marco Braun, Dominic Spata, Matthias Rottmann

TARS, a framework developed by researchers including those from Aptiv Services Deutschland GmbH, introduces a traffic-aware approach to radar scene flow estimation by integrating object detection context with a novel Traffic Vector Field. This method achieves state-of-the-art accuracy, reducing End-Point Error by 29% on the VOD dataset and over 59% on a proprietary dataset, notably improving the inference of tangential motion from sparse radar point clouds.

Key insights

• Capturing rigid motion at a 'traffic-level' within a coarse grid, rather than at an unreliable object-instance level, effectively addresses radar's sparsity for scene flow.
• Leveraging high-level semantic context from an object detection branch significantly boosts the accuracy and robustness of low-level scene flow estimation for sparse sensors.
• Global axial attention within the Traffic Vector Field is crucial for modeling holistic traffic dependencies and accurately inferring challenging tangential motion components from radar data.

Summary by alphaXiv.