Research map / Multi-agent systems and games / Human-AI and goal-conditioned
AGC-Drive: A Large-Scale Dataset for Real-World Aerial-Ground Collaboration in Driving Scenarios
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.
Introduces AGC-Drive, a real-world dataset of two vehicles and one LiDAR-equipped UAV for aerial-ground collaborative 3D perception, and benchmarks six intermediate/late/early collaborative detectors on V2V and V2UAV tasks.
Problem
Collaborative perception research lacks high-quality real-world datasets combining ground vehicles with UAVs, especially UAVs carrying LiDAR in driving scenarios.
Approach
A real-world vehicle-vehicle-UAV dataset with time-synchronized multi-view cameras and LiDAR on all agents, 3D box annotations with occlusion levels, and benchmarks for V2V and vehicle-to-UAV cooperative detection, with ICP-based relative pose estimation.
Key results
- On AGC-V2V, Cobevt achieves the best mAP@0.5 of 46.1% and mAP@0.7 of 41.7% among the evaluated methods.
- On AGC-VUC, adding the UAV improves all intermediate frameworks; V2VNet gains the largest Delta_UAV of +11.5.
- Late fusion PointPillars is lowest on AGC-V2V with 17.7% mAP@0.5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add UAV participation to V2VNet collaboration | V2VNet · AGC-VUC | mAP@0.5: 30.5 -> 40.1 | Helped |
| Add UAV participation to V2VNet collaboration | V2VNet · AGC-VUC | mAP@0.7: 14.6 -> 27.9 | Helped |
| Add UAV participation to Cobevt collaboration | Cobevt · AGC-VUC | mAP@0.5: 42.3 -> 42.9 | Helped |
| Add UAV participation to Cobevt collaboration | Cobevt · AGC-VUC | mAP@0.7: 36.9 -> 37.5 | Helped |
| Add UAV participation to Where2comm collaboration | Where2comm · AGC-VUC | mAP@0.5: 42.6 -> 44.2 | Helped |
| Add UAV participation to Where2comm collaboration | Where2comm · AGC-VUC | mAP@0.7: 30.7 -> 32.0 | Helped |
| Add UAV participation to V2X-ViT collaboration | V2X-ViT · AGC-VUC | mAP@0.5: 38.3 -> 42.6 | Helped |
| Add UAV participation to V2X-ViT collaboration | V2X-ViT · AGC-VUC | mAP@0.7: 28.7 -> 33.9 | Helped |
Surprising findings
- Early fusion is only slightly better than late fusion, and V2VNet collapses at mAP@0.7 (5.7%) on AGC-V2V.
- The UAV's gain is largest for the weakest baseline (V2VNet), and small for Cobevt (+0.6).
Limitations
- UAV LiDAR (32-beam) point clouds are sparse and limit fine-grained object-level assistance.
- Retained time delays and pose errors from ICP-based registration degrade localization accuracy.
- Benchmark results are moderate overall across methods.
Experimental setup
Datasets: AGC-Drive
Benchmarks: AGC-V2V AGC-VUC
Metrics: mAP@0.5 mAP@0.7 Delta_UAV
Baselines: Late Fusion Early Fusion V2VNet Cobevt Where2comm V2X-ViT
Builds on: OPV2V PointPillars nuScenes
Closest papers by meaning
- AirCopBench: A Benchmark for Multi-drone Collaborative Embodied Perception and Reasoning (AAAI 2026 · similarity 0.88)
- STAMP: Scalable Task- And Model-agnostic Collaborative Perception (ICLR 2025 · similarity 0.87)
- UrbanIng-V2X: A Large-Scale Multi-Vehicle, Multi-Infrastructure Dataset Across Multiple Intersections for Cooperative Perception (NeurIPS 2025 · similarity 0.87)
- V2X-Radar: A Multi-modal Dataset with 4D Radar for Cooperative Perception (NeurIPS 2025 · similarity 0.87)
- UAV3D: A Large-scale 3D Perception Benchmark for Unmanned Aerial Vehicles (NeurIPS 2024 · similarity 0.87)
- Transfer Your Perspective: Controllable 3D Generation from Any Viewpoint in a Driving Scene (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.