Research map / 3D vision / Depth and stereo
View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network
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 View-decoupled Transformer (VDT) for aerial-ground person re-ID, using hierarchical subtractive separation of meta/view tokens plus an orthogonal loss, and contributes the synthetic CARGO dataset.
Problem
Person re-identification across heterogeneous aerial and ground cameras suffers from dramatic view discrepancy that disrupts identity representation; prior work focuses on homogeneous camera networks and has few datasets.
Approach
A ViT-based model with an extra meta token (identity) and view token (view); after each self-attention block the view token is subtracted from the meta token, and an orthogonal loss forces the final identity and view tokens to be independent, achieving view-decoupled identity features at the same complexity as ViT.
Key results
- On CARGO, VDT reaches 64.10% Rank1 and 55.20% mAP overall, versus 61.54%/53.54% for ViT.
- On AG-ReID G→A, VDT achieves 86.59% Rank1 and 78.57% mAP, surpassing Explain by 3.74% mAP and 5.18% Rank1.
- VDT adds negligible inference cost: 2.28 ms vs 2.27 ms per image for the ViT baseline on CARGO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use VDT instead of Explain | AG-ReID · AG-ReID G→A | mAP: 73.39 -> 78.57 | Helped |
| Use VDT instead of ViT | CARGO · CARGO A↔G | mAP: 40.11 -> 42.76 | Helped |
| Remove hierarchical subtractive separation (Eq. 4) (ablation) | VDT · CARGO and AG-ReID | performance (Rank1, mAP, and mINP) | Hurt |
| Remove orthogonal loss (Eq. 9) (ablation) | VDT · CARGO and AG-ReID | performance (Rank1, mAP, and mINP) | Hurt |
| Remove both hierarchical subtractive separation and orthogonal loss (ablation) | VDT · CARGO and AG-ReID | performance (Rank1, mAP, and mINP) | Hurt |
Surprising findings
- The optimal loss weight lambda is smaller for datasets with more identities (0.001 on CARGO vs 1.0 on AG-ReID).
Limitations
- View discrepancy is treated as the main challenge; other disturbances such as illumination and occlusion are not explicitly modeled.
- CARGO is synthetic, so transfer to real-world aerial-ground surveillance is not demonstrated in the main text.
Experimental setup
Datasets: CARGO AG-ReID
Benchmarks: CARGO AG-ReID
Metrics: Rank1 mAP mINP
Baselines: SBS PCB BoT MGN VV AGW ViT Explain
Builds on: ViT triplet loss AG-ReID explainable attribute model AGW
Closest papers by meaning
- Semantic-Driven Visual Progressive Refinement for Aerial-Ground Person ReID: A Challenging Large-Scale Benchmark (AAAI 2026 · similarity 0.91)
- AG-VPReID: A Challenging Large-Scale Benchmark for Aerial-Ground Video-based Person Re-Identification (CVPR 2025 · similarity 0.90)
- SeCap: Self-Calibrating and Adaptive Prompts for Cross-view Person Re-Identification in Aerial-Ground Networks (CVPR 2025 · similarity 0.90)
- GSAlign: Geometric and Semantic Alignment Network for Aerial-Ground Person Re-Identification (NeurIPS 2025 · similarity 0.87)
- From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization (CVPR 2025 · similarity 0.83)
- Diffusion-based Synthetic Data Generation for Visible-Infrared Person Re-Identification (AAAI 2025 · similarity 0.83)
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.