Research map / Label-efficient and robust learning / OOD and anomaly detection
Cross-Rejective Open-Set SAR Image Registration
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.
Frames SAR keypoint registration as cross-rejective open-set recognition, using two domain classifiers with cross-domain rejection and supervised contrastive feature-tuning to filter redundant keypoints without post-processing.
Problem
SAR image registration treated as closed-set classification forces redundant pre-detected keypoints (possibly over 90%) into given classes, producing incorrect matched-point pairs.
Approach
Treat redundant keypoints as OOD samples in an open-set formulation where all samples are available during training; two per-image classification domains are trained, cross-domain predictions with statistical rejection generate estimation labels that feed back into a supervised contrastive feature-tuning module iteratively.
Key results
- On YellowR1, CroR-OSIR reaches RMSall 0.2533 with 11 matched pairs, versus 0.3779 for the best prior DALF-based result (DALF 14 pairs) and 0.5487 for STDT-Net.
- On Wuhan, CroR-OSIR achieves RMSall 0.1339 versus 0.4217 for AdaSSIR.
- Removing the cross-domain feature-tuning term at lambda=0.90 raises YellowR1 RMSall from 0.3085 to 1.3697 (ablation).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CroR-OSIR instead of DALF | SAR image registration · YellowR1 | RMSall: 0.3779 -> 0.2533 | Helped |
| use CroR-OSIR instead of STDT-Net | SAR image registration · YellowR2 | RMSall: 0.4808 -> 0.0952 | Helped |
| use CroR-OSIR instead of XFeat | SAR image registration · Yama | RMSall: 0.3295 -> 0.3014 | Helped |
| use CroR-OSIR instead of AdaSSIR | SAR image registration · Wuhan | RMSall: 0.4217 -> 0.1339 | Helped |
| use CroR-OSIR instead of DALF | image registration · Optical images | RMSall: 0.7043 -> 0.6600 | Helped |
| add cross-domain estimation tuning term (L*crd) to SupCon feature-tuning (ablation) | CroR-OSIR · YellowR1, λ=0.90 | RMSall: 1.3697 -> 0.3085 | Helped |
| add cross-domain estimation tuning term (L*crd) to SupCon feature-tuning (ablation) | CroR-OSIR · Yama, λ=0.90 | RMSall: 8.0312 -> 0.4274 | Helped |
| add cross-domain estimation tuning term (L*crd) to SupCon feature-tuning (ablation) | CroR-OSIR · YellowR1, λ=0.95 | RMSall: 0.9104 -> 0.2533 | Helped |
| add cross-domain estimation tuning term (L*crd) to SupCon feature-tuning (ablation) | CroR-OSIR · Yama, λ=0.95 | RMSall: 0.9462 -> 0.3014 | Helped |
| increase rejection threshold λ (ablation) | CroR-OSIR · YellowR1, YellowR2, Yama, and Wuhan | RMSall | Helped |
| set λ to 0.95, 0.97, or 0.99 (ablation) | CroR-OSIR · Four SAR datasets | RMSall | No clear effect |
Surprising findings
- Baseline methods generally need post-processing, while CroR-OSIR outputs matches directly without it.
- Redundant keypoints can exceed 90% of all pre-detected keypoints.
Limitations
- Performance depends on the rejection threshold lambda; too low yields incorrect one-to-many matches, too high prevents finding matches.
- Relies on a well-trained encoder; the iterative procedure requires several warm-up epochs (50 to 500).
Experimental setup
Datasets: Wuhan YellowR1 YellowR2 Yama HPatches
Benchmarks: Wuhan YellowR1 YellowR2 Yama HPatches
Metrics: RMSall RMSLOO Nred
Baselines: SIFT SAR-SIFT DNN SuperPoint Sparse-NCNet MSDF-Net STDT-Net AdaSSIR DALF XFeat DBMDF
Builds on: SupCon OpenMax SIFT Swin Transformer
Closest papers by meaning
- RSAR: Restricted State Angle Resolver and Rotated SAR Benchmark (CVPR 2025 · similarity 0.83)
- SARDet-100K: Towards Open-Source Benchmark and ToolKit for Large-Scale SAR Object Detection (NeurIPS 2024 · similarity 0.82)
- Dual Information Purification for Lightweight SAR Object Detection (AAAI 2025 · similarity 0.80)
- FUSAR-GPT: A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery (CVPR 2026 · similarity 0.80)
- M3LEO: A Multi-Modal, Multi-Label Earth Observation Dataset Integrating Interferometric SAR and Multispectral Data (NeurIPS 2024 · similarity 0.80)
- Learning De-Biased Representations for Remote-Sensing Imagery (NeurIPS 2024 · similarity 0.79)
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.