Abstract:

The Radar dEtection TRansformer (RETR) has recently been introduced to fuse multi-view millimeter-wave radar heatmaps using a detection transformer framework and a simple geometric learning approach for indoor radar perception. A key part of RETR is its tunable positional encoding (TPE), which adjusts the weight of depth positional embeddings across different views to improve feature matching. However, the original design fixes the TPE ratio before training. Differentiable Positional Encoding (DiPE) was proposed to overcome this limitation for bounding box detection by automatically adjusting the TPE ratio with dual differentiable masks on depth and angular positional embeddings. In this paper, we build on the existing DiPE approach and propose a segmentation pipeline that extends its application to human instance segmentation directly from radar signals. Our method integrates the established DiPE mechanism into a framework for segmentation, working with either fixed (e.g., sinusoidal) or learnable positional embeddings, and is optimized end-to-end with a segmentation loss. Evaluation on the open-sourced MMVR dataset shows that our segmentation pipeline achieves improved performance compared to conventional methods.