Research map / Multi-agent systems and games / Human-AI and goal-conditioned
STAMP: Scalable Task- And Model-agnostic Collaborative 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 STAMP, a task- and model-agnostic heterogeneous collaborative perception framework that aligns each agent's BEV features to a shared protocol domain via lightweight adapter-reverter pairs trained with feature and decision alignment losses.
Problem
Heterogeneous agents in collaborative perception differ in sensors, model architectures, and downstream tasks, causing feature domain gaps that hinder efficient and secure cross-agent collaboration, and existing methods scale poorly or share models.
Approach
Train a shared protocol network once, then learn for each agent only a lightweight adapter (local features to protocol domain) and reverter (protocol features back to local domain) with feature-space and decision-space alignment losses, keeping local models, fusion, and decoders private and frozen.
Key results
- On OPV2V with 4 agents at sigma=0.0, STAMP reaches AP@50 0.985 versus 0.983 for HEAL and 0.980 for E2E training.
- In the model- and task-agnostic setting, STAMP raises Agent 2 camera detection AP@50 from 0.399 to 0.760 without noise.
- Training cost grows by about 2.36 GPU hours per additional agent versus 17.07 for existing heterogeneous pipelines (7.2x saving).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add collaborative feature alignment (CFA) to heterogeneous feature fusion (ablation) | STAMP · OPV2V, Agent 2 camera object detection, σ=0.0 | AP@50: 0.399 -> 0.760 | Helped |
| add CFA to heterogeneous feature fusion (ablation) | STAMP · OPV2V, Agent 3 static segmentation, σ=0.0 | MIoU: 0.114 -> 0.624 | Helped |
| add CFA to heterogeneous feature fusion (ablation) | STAMP · OPV2V, Agent 4 dynamic segmentation, σ=0.0 | MIoU: 0.070 -> 0.690 | Helped |
| use a camera-modality protocol (ablation) | STAMP · OPV2V, Agent 2 camera object detection | AP@50: 0.760 -> 0.777 | Helped |
| use a dynamic-segmentation protocol (ablation) | STAMP · OPV2V, Agent 4 dynamic segmentation | MIoU: 0.690 -> 0.723 | Helped |
| use a static-segmentation protocol (ablation) | STAMP · OPV2V, Agent 3 static segmentation | MIoU: 0.624 -> 0.681 | Helped |
| use a static-segmentation protocol (ablation) | STAMP · OPV2V, Agent 4 dynamic segmentation | MIoU: 0.690 -> 0.235 | Hurt |
| replace the CNN-based protocol encoder with Point-transformer (ablation) | STAMP · OPV2V, Agent 1 object detection | AP@50: 0.936 -> 0.942 | Helped |
| reduce protocol BEV feature channels from 64 to 32 or 16 (ablation) | STAMP · OPV2V and V2V4Real | performance | No clear effect |
| replace three ConvNeXt adapter-reverter layers with a single 1×1 convolution (ablation) | STAMP · OPV2V and V2V4Real | performance | No clear effect |
| replace three ConvNeXt adapter-reverter layers with three self-attention layers (ablation) | STAMP · OPV2V and V2V4Real | performance | No clear effect |
| remove decision-space alignment loss L_d (ablation) | STAMP · OPV2V | performance | Hurt |
Surprising findings
- Collaboration degrades the strongest LiDAR agent (Agent 1) in the task-agnostic setting, attributed to a bottleneck from the weaker camera agent.
- Collaboration without feature alignment drops segmentation performance far below the single-agent baseline (e.g., 0.548 to 0.114).
Limitations
- Bottleneck effect: performance of the weakest agent constrains the whole system, degrading strong agents.
- Security analysis is limited; the authors acknowledge the need for further security evaluation.
- Protocol model choice matters: objective mismatch (static segmentation protocol) can severely degrade performance (Agent 4 from 0.690 to 0.235).
Experimental setup
Datasets: OPV2V V2V4Real
Benchmarks: OPV2V V2V4Real
Metrics: AP@30 AP@50 MIoU
Baselines: Late Fusion Calibrator E2E Training HEAL V2X-ViT CoBEVT HM-ViT
Builds on: HEAL Calibrator Pyramid Fusion ConvNeXt
Closest papers by meaning
- CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization (CVPR 2025 · similarity 0.90)
- CP-Guard: Malicious Agent Detection and Defense in Collaborative Bird’s Eye View Perception (AAAI 2025 · similarity 0.90)
- One is Plenty: A Polymorphic Feature Interpreter for Immutable Heterogeneous Collaborative Perception (CVPR 2025 · similarity 0.89)
- An Extensible Framework for Open Heterogeneous Collaborative Perception (ICLR 2024 · similarity 0.88)
- Privacy-Preserving V2X Collaborative Perception Integrating Unknown Collaborators (AAAI 2025 · similarity 0.88)
- AGC-Drive: A Large-Scale Dataset for Real-World Aerial-Ground Collaboration in Driving Scenarios (NeurIPS 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.