Research map / 3D vision / 3D generation
GSAlign: Geometric and Semantic Alignment Network for Aerial-Ground 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.
Proposes GSAlign, a ViT-based aerial-ground person re-identification network with learnable thin plate spline warping (LTPS) and a training-only dynamic alignment module (DAM) that predicts channel-wise visibility masks, evaluated on CARGO, AG-ReID and AG-ReID v2.
Problem
Aerial-ground person re-identification must match pedestrians across extreme viewpoint differences, causing geometric distortion, occlusion, and semantic misalignment between UAV and ground-camera images.
Approach
Insert learnable, rotation-aware thin plate spline warping of patch features into every ViT layer to correct geometric distortion, and use a training-only, input-conditioned channel mask over class prototypes to suppress occluded or noisy dimensions.
Key results
- On CARGO A↔G protocol, GSAlign reaches 64.89% Rank-1, 61.55% mAP, 52.81% mINP, versus VDT's 48.12/42.76/29.95.
- On CARGO ALL protocol, GSAlign achieves 65.06% Rank-1 and 57.95% mAP, improving over VDT by +0.96 Rank-1 and +2.75 mAP.
- On AG-ReID A↔G, GSAlign reaches 83.75% Rank-1 and 75.01% mAP, slightly above VDT (82.91/74.44).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GSAlign instead of VDT | CARGO · A↔G | mAP: 42.76 -> 61.55 (+18.79%) | Helped |
| Use GSAlign instead of VDT | CARGO · ALL | mAP: 55.20 -> 57.95 (+2.75%) | Helped |
| Use GSAlign instead of VDT | AG-ReID · A↔G | mAP: 74.44 -> 75.01 (+0.57%) | Helped |
| Use GSAlign instead of VDT | AG-ReID · G↔A | Rank-1: 86.59 -> 84.10 | Hurt |
| Add LTPS to the baseline (ablation) | VDT on CARGO · A↔G | Rank-1: 48.12 -> 64.89 | Helped |
| Add LTPS to the baseline (ablation) | VDT on CARGO · A↔G | mAP: 42.76 -> 61.08 | Helped |
| Add DAM to Baseline + LTPS (ablation) | GSAlign on CARGO · A↔G | mINP: 50.54% to 52.81% | Helped |
| Use the deformation constraint with LTPS (ablation) | LTPS on CARGO · A↔G | Rank-1: 56.38 -> 65.96 | Helped |
| Use the deformation constraint with LTPS (ablation) | LTPS on CARGO · ALL | Rank-1: 59.94 -> 66.67 | Helped |
| Replace LTPS with original TPS (ablation) | GSAlign on CARGO · G↔G | Rank-1: 83.04 -> 80.36 | Hurt |
| Increase LTPS control points from 4 to 25 (ablation) | GSAlign on CARGO · A↔G | mAP: 61.55 -> 60.23 | Hurt |
| Insert LTPS in all layers instead of only the first layer (ablation) | GSAlign on CARGO · A↔G | Rank-1: 58.51 -> 64.89 | Helped |
Surprising findings
- The original TPS variant (GSAlign-O) underperforms plain ViT on the G↔G protocol, showing standard TPS warping can hurt single-view matching.
- On CARGO A↔A, GSAlign does not improve Rank-1 over VDT (80.00 vs 82.50) and has lower mAP than VDT.
Limitations
- LTPS may fail to learn reliable transforms under extreme viewpoint changes or heavy occlusion.
- LTPS and DAM may slightly hurt aerial-aerial matching where geometric variation is small.
- Evaluated only on specific aerial-ground datasets.
Experimental setup
Datasets: CARGO AG-ReID AG-ReID v2
Benchmarks: CARGO AG-ReID AG-ReID v2
Metrics: Rank-1 mAP mINP
Baselines: SBS PCB BoT MGN VV AGW BAU PAT DTST ViT VDT Explain
Builds on: VDT DPM ViT
Closest papers by meaning
- ProAlign: Progressive Positional and Prototype-Guided Alignment for Aerial-Ground Person Re-Identification (NeurIPS 2026 · similarity 0.95)
- SeCap: Self-Calibrating and Adaptive Prompts for Cross-view Person Re-Identification in Aerial-Ground Networks (CVPR 2025 · similarity 0.90)
- AG-VPReID: A Challenging Large-Scale Benchmark for Aerial-Ground Video-based Person Re-Identification (CVPR 2025 · similarity 0.90)
- View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network (CVPR 2024 · similarity 0.87)
- From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization (CVPR 2025 · similarity 0.83)
- Weakly Misalignment-free Adaptive Feature Alignment for UAVs-based Multimodal Object Detection (CVPR 2024 · 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.