Research map / Label-efficient and robust learning / Recognition and long-tail
SeCap: Self-Calibrating and Adaptive Prompts for Cross-view Person Re-Identification in Aerial-Ground Networks
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 SeCap, a ViT encoder-decoder with a prompt re-calibration module and local feature refinement module for aerial-ground person re-ID, and releases two datasets (LAGPeR, G2APS-ReID).
Problem
Aerial-ground person re-identification suffers from large appearance variation across viewpoints and under-use of local features, with existing attribute-based and view-decoupling methods limited in handling viewpoint diversity.
Approach
Adaptively re-calibrates learnable prompts conditioned on view-invariant global features (Prompt Re-calibration Module) and uses them in a two-way attention decoder to extract view-invariant local features, building on the VDT encoder.
Key results
- On LAGPeR A→G, SeCap reaches 41.79 Rank-1 / 30.37 mAP, versus 40.15 / 28.97 for VDT.
- On G2APS-ReID G→A, SeCap reaches 75.31 Rank-1 / 58.57 mAP, above VDT's 73.05 / 56.23 for the A→G setting... with G→A Rank-1 24.39 vs 19.50 for VDT.
- On AG-ReID.v1 A→G, SeCap reaches 84.03 Rank-1 / 76.16 mAP, above VDT's 82.91 / 74.44.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SeCap instead of VDT | ViT-based AGPReID · LAGPeR A→G | Rank-1: 40.15 -> 41.79 | Helped |
| Use SeCap instead of MIP | ViT-based AGPReID · LAGPeR G→A | Rank-1: 33.90 -> 35.26 | Helped |
| Use SeCap instead of VDT | ViT-based AGPReID · G2APS-ReID A→G | mAP: 56.23 -> 58.57 | Helped |
| Use SeCap instead of VDT | ViT-based AGPReID · AG-ReID.v1 G→A | mAP: 78.57 -> 78.34 | Hurt |
| Add VDT to the ViT baseline (ablation) | ViT-based ReID · LAGPeR A→G | Rank-1: 38.67 -> 40.15 | Helped |
| Add LFRM to the ViT baseline (ablation) | ViT-based ReID · LAGPeR A→G | Rank-1: 38.67 -> 40.05 | Helped |
| Add LFRM and PRM without VDT (ablation) | ViT-based ReID · LAGPeR A→G | Rank-1: 38.67 -> 39.36 | Helped |
| Add LFRM and VDT without PRM (ablation) | ViT-based ReID · LAGPeR A→G | Rank-1: 38.67 -> 41.56 | Helped |
| Add PRM to LFRM and VDT (ablation) | ViT-based ReID · LAGPeR A→G | Rank-1: 41.56 -> 41.79 | Helped |
| Add PRM to LFRM without VDT (ablation) | ViT-based ReID · LAGPeR A→G | Rank-1: 40.05 -> 39.36 | Hurt |
| Add LFRM to VDT (ablation) | ViT-based ReID · LAGPeR A→G | Rank-1: 40.15 -> 41.56 | Helped |
Surprising findings
- Adding PRM without VDT decoupling degrades performance (#4 is no better than LFRM alone), which the authors attribute to prompts learning incorrect view-invariant information.
Limitations
- Gains over VDT are modest on several settings (e.g., AG-ReID.v1 G→A is comparable, not better).
- Absolute performance on LAGPeR remains low (around 30-41 Rank-1 in A→G).
Experimental setup
Datasets: LAGPeR G2APS-ReID AG-ReID.v1 AG-ReID.v2 CARGO
Benchmarks: LAGPeR G2APS-ReID AG-ReID.v1
Metrics: Rank-1 mAP
Baselines: ViT TransReID CLIP-ReID MIP AG-ReID VDT
Builds on: VDT TransReID CLIP-ReID AG-ReID
Closest papers by meaning
- AG-VPReID: A Challenging Large-Scale Benchmark for Aerial-Ground Video-based Person Re-Identification (CVPR 2025 · similarity 0.91)
- GSAlign: Geometric and Semantic Alignment Network for Aerial-Ground Person Re-Identification (NeurIPS 2025 · similarity 0.90)
- View-decoupled Transformer for Person Re-identification under Aerial-ground Camera Network (CVPR 2024 · similarity 0.90)
- Cheb-GR: Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification (CVPR 2025 · similarity 0.82)
- From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization (CVPR 2025 · similarity 0.82)
- Multi-Modal Aerial-Ground Cross-View Place Recognition with Neural ODEs (CVPR 2025 · similarity 0.82)
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.