Research map / Detection and segmentation / Weakly supervised detection and segmentation
RSAR: Restricted State Angle Resolver and Rotated SAR Benchmark
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 Unit Cycle Resolver (UCR), a unit-circle constraint loss on angle encodings that improves weakly supervised rotated detection (H2RBox-v2), and uses it to annotate RSAR, a large multi-class rotated SAR detection dataset.
Problem
Angle boundary discontinuity in rotated object detection makes angle prediction inaccurate, especially for weakly supervised (horizontal-box-only) models, which hinders building a large rotated SAR dataset.
Approach
Unifies angle resolvers (PSC, ACM) as dimensional mappings and shows they ignore the unit-circle constraint on encoding states; adds a unit cycle constraint loss plus an invalid central region so predicted (cos, sin)-style encodings stay on the unit circle.
Key results
- On DOTA-v1.0, UCR with H2RBox-v2 improves AP75 from 24.07 (PSC) to 26.17 on RSAR and reaches weakly supervised AP50 of 73.99 on DOTA-v1.0 (vs 72.31 PSC baseline), surpassing fully supervised FCOS (AP50 71.44).
- On DOTA-v1.0, UCR (3-dim) reaches AP75 of 42.10 vs 39.49 for PSC baseline, and mAP 43.10 vs 40.69.
- RSAR contains 95,842 images and 183,534 instances across 6 categories, described as the largest multi-class rotated SAR detection dataset.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UCR with 3D mapping instead of PSC | H2RBox-v2 · RSAR | AP75: 24.07 -> 26.17 | Helped |
| Use UCR with 3D mapping instead of PSC | H2RBox-v2 · DOTA-v1.0 | AP75: 39.49 -> 42.10 | Helped |
| Use UCR with 3D mapping instead of PSC | H2RBox-v2 · HRSC | AP75: 64.80 -> 75.49 | Helped |
| Add unit cycle loss with λuc=0.05 instead of removing it (ablation) | H2RBox-v2 + UCR (3D) · RSAR | AP75: 24.07 -> 26.17 | Helped |
| Set unit cycle loss weight to 0.01 instead of removing the loss (ablation) | H2RBox-v2 + UCR (3D) · RSAR | AP75: 24.07 -> 24.10 | No clear effect |
| Set unit cycle loss weight to 0.03 instead of removing the loss (ablation) | H2RBox-v2 + UCR (3D) · RSAR | AP75: 24.07 -> 23.67 | Hurt |
| Set unit cycle loss weight to 0.1 instead of removing the loss (ablation) | H2RBox-v2 + UCR (3D) · RSAR | AP75: 24.07 -> 25.86 | Helped |
| Set unit cycle loss weight to 0.2 instead of removing the loss (ablation) | H2RBox-v2 + UCR (3D) · RSAR | AP75: 24.07 -> 25.10 | Helped |
| Add unit cycle loss at λuc=0.03 instead of removing it (ablation) | H2RBox-v2 + UCR (3D) · DOTA-v1.0 | AP75: 39.49 -> 42.10 | Helped |
| Add the unit circle loss (ablation) | UCR with 2D mapping · RSAR | AP75: 40.17 -> 41.57 | Helped |
| Add the invalid region (ablation) | UCR with 2D mapping · RSAR | AP75: 40.17 -> 40.94 | Helped |
| Combine unit circle loss with the invalid region (ablation) | UCR with 2D mapping · RSAR | AP75: 40.17 -> 42.26 | Helped |
Surprising findings
- Weakly supervised H2RBox-v2 with UCR surpasses fully supervised FCOS on DOTA-v1.0 in AP50 (73.99 vs 71.44) and mAP.
Limitations
- Designing a fully supervised detector tailored to SAR with RSAR is left for future work.
- Ablation on loss weight shows performance declines for larger unit-cycle loss weights.
- Some aircraft orientation data was excluded due to ambiguity.
Experimental setup
Datasets: RSAR SARDet-100K DOTA-v1.0 HRSC
Benchmarks: RSAR DOTA-v1.0 HRSC
Metrics: AP50 AP75 mAP
Baselines: H2RBox H2RBox-v2 with ACM H2RBox-v2 with PSC FCOS RetinaNet R3Det S2ANet Faster RCNN O-RCNN ReDet RoI-Transformer Deformable DETR
Builds on: H2RBox-v2 PSC ACM FCOS SARDet-100K
Closest papers by meaning
- SARDet-100K: Towards Open-Source Benchmark and ToolKit for Large-Scale SAR Object Detection (NeurIPS 2024 · similarity 0.86)
- Learning De-Biased Representations for Remote-Sensing Imagery (NeurIPS 2024 · similarity 0.84)
- Cross-Rejective Open-Set SAR Image Registration (CVPR 2025 · similarity 0.83)
- Gamma Distribution PCA-Enhanced Feature Learning for Angle-Robust SAR Target Recognition (ICML 2025 · similarity 0.83)
- ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects (CVPR 2025 · similarity 0.82)
- FUSAR-GPT: A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery (CVPR 2026 · 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.