Research map / Detection and segmentation / CNN features
Theoretically Achieving Continuous Representation of Oriented Bounding Boxes
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 COBB, a 9-parameter continuous oriented bounding box representation (outer HBB, sliding ratio, four IoU scores) that is theoretically continuous and plugs into existing oriented detectors, evaluated on a new JDet-based benchmark.
Problem
Oriented bounding box regression suffers from encoding discontinuity (rotation and aspect ratio) and decoding discontinuity (incompleteness and ambiguity), which hurts precision of oriented object detectors.
Approach
Represent an OBB by its outer horizontal box, a continuous sliding ratio derived from the OBB area ratio, and four IoU scores that disambiguate the four OBBs sharing the same outer box and sliding ratio; this gives a provably continuous encoding and decoding that can replace the OBB representation in existing detectors as a plugin.
Key results
- On DOTA-v1.0 with Faster R-CNN, COBB-sig improves mAP75 from 40.13 to 44.03 (+3.9 points) and COBB-ln reaches 44.08 mAP75.
- On HRSC2016, COBB-ln outperforms Gliding Vertex by 13.77% in mAP75.
- With RoI Transformer as baseline on DOTA-v1.0, COBB-ln-ln reaches 76.53 mAP50 and 50.41 mAP75 versus 75.59 and 48.54 for the baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IoU scores instead of one-hot coding for OBB classification (ablation) | Rotated Faster R-CNN + COBB-sig · DOTA-v1.0 | mAP50: 73.46 -> 74.00 | Helped |
| Use IoU scores instead of one-hot coding for OBB classification (ablation) | Rotated Faster R-CNN + COBB-sig · DOTA-v1.0 | mAP75: 43.76 -> 44.03 | Helped |
| Use COBB-ln on horizontal proposals (ablation) | RoI Transformer · DOTA-v1.0 | mAP50: 75.59 -> 76.27 | Helped |
| Use COBB-ln on horizontal proposals (ablation) | RoI Transformer · DOTA-v1.0 | mAP75: 48.54 -> 50.23 | Helped |
| Use COBB-ln on oriented proposals (ablation) | RoI Transformer · DOTA-v1.0 | mAP50: 75.59 -> 76.10 | Helped |
| Use COBB-ln on oriented proposals (ablation) | RoI Transformer · DOTA-v1.0 | mAP75: 48.54 -> 48.11 | Hurt |
| Use COBB-ln on both horizontal and oriented proposals (ablation) | RoI Transformer · DOTA-v1.0 | mAP50: 75.59 -> 76.53 | Helped |
| Use sliding ratio rs rather than area ratio ra as the regression target (ablation) | Rotated Faster R-CNN + COBB-ln · DOTA-v1.0 | mAP50: 74.13 -> 74.44 | Helped |
| Use sliding ratio rs rather than area ratio ra as the regression target (ablation) | Rotated Faster R-CNN + COBB-ln · DOTA-v1.0 | mAP75: 43.31 -> 44.08 | Helped |
| Replace Rotated Faster R-CNN OBB representation with COBB-ln | Rotated Faster R-CNN · DOTA-v1.0 | mAP75: 40.13 -> 44.08 | Helped |
| Replace Rotated Faster R-CNN OBB representation with COBB-ln | Rotated Faster R-CNN · HRSC2016 | mAP75: 31.64 -> 72.29 | Helped |
Surprising findings
- One-hot classification of the four OBB styles performs worse than IoU-score targets, attributed to the discontinuity one-hot coding introduces.
- Gliding Vertex outperforms Faster R-CNN only marginally in mAP50 on DOTA-v1.0 (0.30%) despite its claimed advantages.
Limitations
- COBB's outer HBB, sliding ratio and IoU scores vary irregularly under rotation, limiting use with rotation-equivariant detectors such as ReDet.
- Gains on DOTA-v1.5 and mAP50 are modest (around 1 point), and mAP50 improvements for some baselines are small.
Experimental setup
Datasets: DOTA-v1.0 DOTA-v1.5 DIOR HRSC2016 FAIR1M-1.0 FAIR1M-2.0
Benchmarks: DOTA-v1.0 DOTA-v1.5 DIOR HRSC2016 FAIR1M-1.0 FAIR1M-2.0
Metrics: mAP50 mAP75 mAP50:95
Baselines: Rotated Faster R-CNN Gliding Vertex RoI Transformer Oriented R-CNN KLD ReDet
Builds on: Faster R-CNN Rotated Faster R-CNN Gliding Vertex CSL GWD
Closest papers by meaning
- Rethinking Boundary Discontinuity Problem for Oriented Object Detection (CVPR 2024 · similarity 0.87)
- ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects (CVPR 2025 · similarity 0.86)
- Bounding Box Stability against Feature Dropout Reflects Detector Generalization across Environments (ICLR 2024 · similarity 0.86)
- ReDiffDet: Rotation-equivariant Diffusion Model for Oriented Object Detection (CVPR 2025 · similarity 0.85)
- Towards Flexible 3D Perception: Object-Centric Occupancy Completion Augments 3D Object Detection (NeurIPS 2024 · similarity 0.85)
- OpenBox: Annotate Any Bounding Boxes in 3D (NeurIPS 2025 · similarity 0.85)
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.