Research map / Video / Video segmentation and tracking
AG-VPReID: A Challenging Large-Scale Benchmark for Aerial-Ground Video-based Person Re-Identification
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 AG-VPReID, a large-scale aerial-ground video person ReID dataset (6,632 subjects, 32,321 tracklets, 9.6M frames), and AG-VPReID-Net, a three-stream model evaluated against video ReID baselines.
Problem
Video-based person re-identification across aerial (drone) and ground (CCTV, wearable) cameras suffers from viewpoint, scale, and resolution gaps, and existing datasets are small or single-platform.
Approach
A new multi-platform aerial-ground video ReID dataset with drones at 15-120m altitude, plus a three-stream network combining temporal 3D shape modeling, physics-informed UV-texture appearance normalization, and multi-scale frozen CLIP features with a transformer decoder.
Key results
- AG-VPReID-Net reaches 64.0% mAP and 71.9% Rank-1 on aerial-to-ground AG-VPReID, and 58.0% mAP / 75.6% Rank-1 on ground-to-aerial.
- On G2A-VReID it achieves 81.3% mAP and 73.1% Rank-1, surpassing MGH by 4.6% mAP.
- On MARS it reports 91.5% mAP and 93.2% Rank-1, outperforming CLIP-ReID by 3.4% mAP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the Adapted Temporal-Spatial stream (ablation) | AG-VPReID-Net · AG-VPReID Aerial-to-Ground | Rank-1: 65.70 -> 71.91 | Helped |
| Add the Adapted Temporal-Spatial stream (ablation) | AG-VPReID-Net · AG-VPReID Ground-to-Aerial | Rank-1: 70.10 -> 75.57 | Helped |
| Add the Normalized Appearance stream (ablation) | AG-VPReID-Net · AG-VPReID Aerial-to-Ground | Rank-1: 71.80 -> 71.91 | Helped |
| Add the Normalized Appearance stream (ablation) | AG-VPReID-Net · AG-VPReID Ground-to-Aerial | Rank-1: 75.40 -> 75.57 | Helped |
| Add the Multi-Scale Attention stream (ablation) | AG-VPReID-Net · AG-VPReID Aerial-to-Ground | Rank-1: 69.50 -> 71.91 | Helped |
| Add the Multi-Scale Attention stream (ablation) | AG-VPReID-Net · AG-VPReID Ground-to-Aerial | Rank-1: 72.80 -> 75.57 | Helped |
| Use all three streams instead of only the Adapted Temporal-Spatial stream (ablation) | AG-VPReID-Net · AG-VPReID Aerial-to-Ground | Rank-1: 71.52 -> 71.91 | Helped |
| Use all three streams instead of only the Adapted Temporal-Spatial stream (ablation) | AG-VPReID-Net · AG-VPReID Ground-to-Aerial | Rank-1: 74.80 -> 75.57 | Helped |
| Increase aerial altitude from 15m to 120m (ablation) | AG-VPReID-Net · AG-VPReID Aerial-to-Ground | Rank-1: 80.66 -> 53.00 | Hurt |
| Increase aerial altitude from 15m to 120m (ablation) | St-1 · AG-VPReID Aerial-to-Ground | Rank-1: 80.28 -> 52.47 | Hurt |
| Increase aerial altitude from 15m to 120m (ablation) | St-2 · AG-VPReID Aerial-to-Ground | Rank-1: 69.75 -> 38.12 | Hurt |
| Increase aerial altitude from 15m to 120m (ablation) | St-3 · AG-VPReID Aerial-to-Ground | Rank-1: 74.25 -> 35.13 | Hurt |
Surprising findings
- Performance drops sharply with drone altitude; Rank-1 falls by 27.66% from 15m to 120m, more than clothing changes cause (7.85% in ground-to-ground).
Limitations
- Performance across all methods remains low on AG-VPReID, indicating the benchmark is hard but also that current methods struggle.
- Gains of the proposed streams are described as incremental.
- Ablation and results rely on the authors' split of 3,013 identities.
Experimental setup
Datasets: AG-VPReID iLIDS-VID MARS LS-VID G2A-VReID
Benchmarks: AG-VPReID MARS LS-VID iLIDS-VID G2A-VReID
Metrics: mAP CMC Rank-1 Rank-5 Rank-10
Baselines: CLIP-ReID LSTRL MGH STMP M3D GLTR TCLNet CAVIT DCCT MFA
Builds on: CLIP-ReID PhysPT Texformer CLIP TMT
Closest papers by meaning
- SeCap: Self-Calibrating and Adaptive Prompts for Cross-view Person Re-Identification in Aerial-Ground Networks (CVPR 2025 · similarity 0.91)
- View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network (CVPR 2024 · similarity 0.90)
- GSAlign: Geometric and Semantic Alignment Network for Aerial-Ground Person Re-Identification (NeurIPS 2025 · similarity 0.90)
- Exploring the Camera Bias of Person Re-identification (ICLR 2025 · similarity 0.86)
- Alice Benchmarks: Connecting Real World Re-Identification with the Synthetic (ICLR 2024 · similarity 0.85)
- NightReID: A Large-Scale Nighttime Person Re-Identification Benchmark (AAAI 2025 · similarity 0.85)
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.