Research map / Multi-agent systems and games / Human-AI and goal-conditioned
What Makes Good Collaborative Views? Contrastive Mutual Information Maximization for Multi-Agent Perception
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 CMiMC, which maximizes multi-view mutual information between pre- and post-collaboration BEV features via contrastive learning (CMiMNet) to train an intermediate collaboration encoder for LiDAR 3D object detection.
Problem
Intermediate collaboration in multi-agent perception treats feature fusion as an opaque process trained only on downstream loss, risking loss of discriminative information from individual views.
Approach
Defines multi-view mutual information (MVMI) between one collaborative view and many individual views on global and local scales, and estimates/maximizes it with a JS-divergence contrastive discriminator (CMiMNet) during training only, added to the detection loss.
Key results
- CMiMC achieves 61.58% AP@0.5 and 55.80% AP@0.7 on V2X-Sim 1.0, improving over DiscoNet by 3.08% and 4.44%.
- CMiMC with 1/32 compression ratio achieves 59.78% AP@0.5 with communication volume of 32.8 KB versus 1.05e3 KB for uncompressed intermediate methods, still above DiscoNet's SOTA AP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CMiMC instead of DiscoNet | intermediate collaboration for 3D object detection · V2X-Sim 1.0 | AP at IoU 0.5: 3.08% | Helped |
| use CMiMC instead of DiscoNet | intermediate collaboration for 3D object detection · V2X-Sim 1.0 | AP at IoU 0.7: 4.44% | Helped |
| add knowledge distillation (ablation) | DiscoNet · V2X-Sim 1.0 | AP at IoU 0.5: 1.41% | Helped |
| add knowledge distillation (ablation) | DiscoNet · V2X-Sim 1.0 | AP at IoU 0.7: 0.19% | Helped |
| compress intermediate features at ratio 1/32 (ablation) | CMiMC · V2X-Sim 1.0 | AP at IoU 0.5: 61.58 -> 59.78 | Hurt |
| compress intermediate features at ratio 1/32 (ablation) | CMiMC · V2X-Sim 1.0 | AP at IoU 0.7: 55.80 -> 54.04 | Hurt |
| compress intermediate features at ratio 1/32 (ablation) | CMiMC · V2X-Sim 1.0 | communication volume: $1.05\times 10^3$ -> $3.28\times 10^1$ | Helped |
Surprising findings
- Using knowledge distillation from early collaboration improves DiscoNet by only 1.41% and 0.19% AP, suggesting the early-collaboration teacher is a bottleneck.
Limitations
- Evaluated only on one simulated dataset in the main results (other datasets only in appendix).
- MI estimation requires training-time negative pairs from different scenes, adding training cost.
Experimental setup
Datasets: V2X-Sim 1.0
Benchmarks: V2X-Sim 1.0
Metrics: average precision at IoU 0.5 average precision at IoU 0.7 communication volume (KB) AP@0.5 under localization noise
Baselines: No Collaboration Early Collaboration Late Collaboration Who2com When2com V2VNet Where2comm DiscoNet V2X-ViT
Builds on: DiscoNet Deep InfoMax Contrastive Multiview Coding MINE InfoNCE
Closest papers by meaning
- An Extensible Framework for Open Heterogeneous Collaborative Perception (ICLR 2024 · similarity 0.89)
- Communication-Efficient Collaborative Perception via Information Filling with Codebook (CVPR 2024 · similarity 0.88)
- NegoCollab: A Common Representation Negotiation Approach for Heterogeneous Collaborative Perception (NeurIPS 2025 · similarity 0.88)
- Pragmatic Heterogeneous Collaborative Perception via Generative Communication Mechanism (NeurIPS 2025 · similarity 0.87)
- CoPEFT: Fast Adaptation Framework for Multi-Agent Collaborative Perception with Parameter-Efficient Fine-Tuning (AAAI 2025 · similarity 0.87)
- CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization (CVPR 2025 · similarity 0.86)
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.