Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Communication-Efficient Collaborative Perception via Information Filling with Codebook
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 CodeFilling, a communication-efficient collaborative 3D detection system that sends integer codebook indices instead of feature maps and selects messages via information-filling optimization.
Problem
Collaborative perception faces a trade-off between detection performance and communication cost, since intermediate collaboration requires transmitting high-dimensional feature maps and suffers from redundant messages across multiple agents.
Approach
Combine a task-driven shared codebook (transmit integer code indices instead of float features, channel-wise compression) with an information-filling selection that collectively chooses non-redundant spatial regions across supporters up to a demand bound (spatial-wise compression).
Key results
- CodeFilling achieves 1333x and 1206x lower communication volume than Where2comm on DAIR-V2X and OPV2VH+ while outperforming it.
- Improves SOTA by 11.093/5.271/38.357% for LiDAR/camera/heterogeneous on DAIR-V2X and OPV2VH+ at a bandwidth constrained by 100K.
- Code index representation reduces communication cost by 208x with the same detection performance; information-filling selection gives 4.8% higher detection performance at the same cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use code-index representation instead of transmitting feature vectors (ablation) | CodeFilling · Ablation setting in Fig. 9a | communication cost: 208 times | Helped |
| use code-index representation instead of transmitting feature vectors (ablation) | CodeFilling · Ablation setting in Fig. 9a | detection performance | No clear effect |
| add information-filling-driven message selection (ablation) | CodeFilling · Ablation setting in Fig. 9a | detection performance: +4.8% | Helped |
| use sum-based utility rather than max utility or no selection (ablation) | CodeFilling information-filling optimization · Fig. 9b ablation | perception-communication trade-off | Helped |
| lower the information demand u (ablation) | CodeFilling · Fig. 9c ablation | communication efficiency under limited budgets | Helped |
| raise the information demand u (ablation) | CodeFilling · Fig. 9c ablation | detection performance under ample budgets | Helped |
| increase codebook size (ablation) | CodeFilling · Fig. 9d ablation | detection performance | Helped |
| decrease codebook size (ablation) | CodeFilling · Fig. 9d ablation | communication efficiency | Helped |
| increase the number of codes used to represent each input vector (ablation) | CodeFilling · Fig. 9e ablation | detection performance | Helped |
| decrease the number of codes used to represent each input vector (ablation) | CodeFilling · Fig. 9e ablation | communication efficiency | Helped |
| use CodeFilling rather than Where2comm | Collaborative 3D detection · DAIR-V2X and OPV2VH+ | communication volume: 1333/115/863, 1206/1078/252 times less | Helped |
| use CodeFilling rather than previous SOTA under bandwidth constrained by a factor of 100K | Collaborative 3D detection · DAIR-V2X and OPV2VH+; LiDAR, camera, and heterogeneous | detection performance: 11.093/5.271/38.357%, 14.75/28.516/28.372% | Helped |
Surprising findings
- Sum-based selection utility outperforms max-based and no-selection, indicating that even top-performing agents benefit from collaboration.
Limitations
- Temporal dimension not explored (stated as future work)
- Information demand upper bound u and codebook configuration are hyperparameters requiring tuning
Experimental setup
Datasets: DAIR-V2X OPV2VH+
Benchmarks: DAIR-V2X OPV2VH+
Metrics: AP@0.30 AP@0.50 communication volume (log2 bytes)
Baselines: No Collaboration Where2comm HMViT V2VNet DiscoNet V2X-ViT AttFuse Late fusion
Builds on: Where2comm When2com CaDDN PointPillar OpenCDA
Closest papers by meaning
- What Makes Good Collaborative Views? Contrastive Mutual Information Maximization for Multi-Agent Perception (AAAI 2024 · similarity 0.88)
- An Extensible Framework for Open Heterogeneous Collaborative Perception (ICLR 2024 · similarity 0.88)
- Pragmatic Heterogeneous Collaborative Perception via Generative Communication Mechanism (NeurIPS 2025 · similarity 0.88)
- NegoCollab: A Common Representation Negotiation Approach for Heterogeneous Collaborative Perception (NeurIPS 2025 · similarity 0.88)
- 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 0.87)
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.