Reproduce “TransRAD: Retentive Vision Transformer for Enhanced Radar Object Detection” on OpenResearch

About this paper

Lei Cheng, Siyang Cao

A Retentive Vision Transformer architecture, TransRAD, achieves state-of-the-art 3D radar object detection by leveraging a specialized attention mechanism and tailored post-processing. The model demonstrates a 9% improvement in AP 0.3 to 61.90% over previous baselines on the RADDet dataset, while simultaneously achieving 4.37 ms inference time and a compact 5.78 million parameters.

Key insights

• The Retentive Manhattan Self-Attention (MaSA) mechanism, extended from RetNet, efficiently captures global context while focusing on salient local features, directly addressing the unique spatial distribution of radar targets.
• Employing an anchor-free, decoupled head architecture effectively resolves conflicts between objectness, classification, and localization tasks, crucial for radar data's unreliable classification.
• A comprehensive loss function, including 2D bounding box and center losses alongside a simple 3D IoU loss, combined with an FPN neck, enables robust detection of small and irregularly shaped objects across 3D dimensions using 2D feature maps.

Summary by alphaXiv.