Research map / Multi-agent systems and games / Human-AI and goal-conditioned
CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CoSDH, a communication-efficient collaborative 3D object detection framework that uses supply-demand-aware region selection, multi-scale feature compression, and confidence-aware late fusion on top of intermediate fusion.
Problem
Multi-agent collaborative perception for autonomous driving faces a trade-off between communication bandwidth and detection accuracy; existing intermediate methods degrade sharply under real V2X bandwidth limits (about 27 Mbps shared across agents).
Approach
Refines collaboration region selection by modeling supply (collaborator's foreground confidence) and demand (ego's low point-cloud density) so that well-observed background areas are not requested, and adds a confidence-aware late fusion stage that compensates intermediate fusion results at very low bandwidth cost.
Key results
- On OPV2V under the 6.75 Mbps per-agent real-world limit, CoSDH reaches 92.92% AP@0.7 at 2.0 Mbps, versus 89.86% for Where2comm at 5.4 Mbps.
- At default settings CoSDH uses 1.1 Mbps on V2XSim (AP@0.7 86.31%) versus 27.5 Mbps for Where2comm at 83.66%.
- CoSDH reports AP@0.7 gains of 3.06%/2.34%/3.75% over the previous best on OPV2V/V2XSim/DAIR-V2X under real-world bandwidth limits.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoSDH rather than CoAlign | PointPillar collaborative perception · OPV2V, BD ≤6.75 Mbps | AP@0.7: 89.86% -> 92.92% | Helped |
| Use CoSDH rather than CoAlign | PointPillar collaborative perception · V2XSim, BD ≤6.75 Mbps | AP@0.7: 83.97% -> 86.31% | Helped |
| Use CoSDH rather than CoAlign | PointPillar collaborative perception · DAIR-V2X, BD ≤6.75 Mbps | AP@0.7: 60.19% -> 63.76% | Helped |
| Add confidence-aware late fusion (ablation) | CoSDH · OPV2V | AP@0.7: 92.60% -> 92.92% | Helped |
| Add demand-mask selection (ablation) | CoSDH with supply selection · OPV2V | AP@0.7: 92.62% -> 92.60% | No clear effect |
| Add supply-mask selection (ablation) | CoSDH with compression and FP16 · OPV2V | AP@0.7: 92.99% -> 92.62% | Hurt |
| Add autoencoder compression (ablation) | CoSDH · OPV2V | AP@0.7: 92.62% -> 92.99% | Helped |
| Convert transmitted features to FP16 (ablation) | CoSDH with autoencoder compression · OPV2V | AP@0.7: 92.99% -> 92.99% | No clear effect |
| Add confidence-aware late fusion (ablation) | CoSDH · OPV2V, intermediate bandwidth 0.05 Mbps | AP@0.7: +4% | Helped |
Surprising findings
- On DAIR-V2X, naive late fusion performs worse than no fusion because it merges many low-quality detections from other agents.
- Under bandwidth limits, some methods improve after autoencoder compression, which the authors attribute to added parameters rather than communication effects.
Limitations
- Evaluated only on LiDAR-based 3D detection with PointPillar backbone.
- Late fusion hyperparameters (confidence threshold, suppression coefficient) are tuned per dataset, with different values for DAIR-V2X.
Experimental setup
Datasets: OPV2V V2XSim DAIR-V2X
Benchmarks: OPV2V V2XSim DAIR-V2X
Metrics: AP@0.5 AP@0.7 bandwidth (Mbps)
Baselines: No Fusion Early Fusion Late Fusion When2com FCooper AttFuse V2VNet DiscoNet V2X-ViT Where2comm CoAlign
Builds on: Where2comm OpenCOOD PointPillar CoAlign
Closest papers by meaning
- STAMP: Scalable Task- And Model-agnostic Collaborative Perception (ICLR 2025 · similarity 0.90)
- An Extensible Framework for Open Heterogeneous Collaborative Perception (ICLR 2024 · similarity 0.89)
- RCP-Bench: Benchmarking Robustness for Collaborative Perception Under Diverse Corruptions (CVPR 2025 · similarity 0.89)
- Multi-agent Collaborative Perception via Motion-aware Robust Communication Network (CVPR 2024 · similarity 0.88)
- CoPEFT: Fast Adaptation Framework for Multi-Agent Collaborative Perception with Parameter-Efficient Fine-Tuning (AAAI 2025 · similarity 0.87)
- CP-Guard: Malicious Agent Detection and Defense in Collaborative Bird’s Eye View Perception (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.