Reproduce “CoDS: Robust Collaborative Perception via Expert-driven Detection and BEV Segmentation” on OpenResearch

About this paper

Jinlong Wang, Yuang Jia, Junhong Lin, Nannan Li, Wei Gao

Researchers developed CoDS, a collaborative perception framework that jointly optimizes 3D object detection and BEV semantic segmentation by mitigating multi-source noise through a Collaborative Reliability Map, semantic mixture-of-experts, and bidirectional task interaction. The framework achieves 4.70, 4.64, and 3.23 point improvements in [email protected], [email protected], and [email protected] respectively, and a 5.27 point gain in Dynamic IoU on the OPV2V dataset, demonstrating enhanced robustness against localization errors and communication latency.

Key insights

• Explicitly quantifying the spatial inconsistency of collaborative fusion quality (via CoRM) provides a crucial prior for robustly processing multi-agent features.
• Differentiating feature extraction with specialized experts (S-MoE) based on semantic content and fusion reliability effectively addresses the varying feature requirements of detection (fine-grained) and segmentation (global context).
• Bidirectional, reliability-gated interaction between 3D object detection and BEV semantic segmentation can mutually enhance their robustness and accuracy by leveraging spatial priors and boundary cues in noisy collaborative environments.

Summary by alphaXiv.