Research map / Detection and segmentation / CNN features
Rethinking Boundary Discontinuity Problem for Oriented Object Detection
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.
Shows that IoU-like smoothing losses do not truly fix boundary discontinuity in oriented object detection, and proposes ACM, a dual-optimization paradigm that predicts a complex-exponential angle encoding (explicitly supervised) and decodes it for joint IoU-like losses.
Problem
Boundary discontinuity problem in oriented object detection: angle predictions become unstable and mutate when objects rotate near the angular boundary, and it was believed IoU-like joint-optimization losses solve it.
Approach
Decouples reversibility (angle correction via a continuous, differentiable complex-exponential encoding exp(j*2*theta) with explicit supervision) from joint optimization (IoU-like loss applied on the decoded angle), so the detector regresses a smooth angular encoding rather than the discontinuous angle itself.
Key results
- On HRSC2016 (Ship) with CenterNet, ACM-KFIoU raises AP75 from 62.95 to 87.77 (+24.82) and ACM-GWD from 61.87 to 86.71.
- On UCAS-AOD (Plane), ACM-KFIoU improves AP75 from 16.81 to 74.48 (+57.67), showing gains for square-like objects.
- On DOTA-v1.0 with CenterNet-ACM, AP50 reaches 78.53; RoI-Trans.-ACM reaches 79.45 AP50, competitive with listed SOTA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ACM to joint-optimization losses | GWD, KLD, KFIoU, and SkewIoU with CenterNet · HRSC2016 (Ship), UCAS-AOD (Car), UCAS-AOD (Plane), and DOTA-v1.0 | AP75: GWD: 61.87 -> 86.71; 28.46 -> 29.15; 38.22 -> 76.00; 34.98 -> 41.97. KLD: 79.29 -> 87.45; 29.99 -> 30.40; 29.19 -> 75.65 | Helped |
| Add ACM with CenterNet or RoI-Transformer | Oriented detectors · DOTA-v1.0 | AP50: 78.32 -> 78.53; 78.07 -> 79.45 | Helped |
| Use a longer encoding length for CSL (ablation) | CSL on HRSC2016 · HRSC2016 | AP75: 2.56 -> 13.58 -> 61.43 -> 77.76 | Helped |
| Vary PSC encoding length (ablation) | PSC on HRSC2016 · HRSC2016 | AP75: 79.20; 79.54; 79.86; 79.51 | No clear effect |
| Use ACM encoding length 2 rather than direct angle prediction (ablation) | CenterNet on HRSC2016 · HRSC2016 | AP75: 62.95 -> 86.33 | Helped |
| Use implicit rather than explicit angle encoding (ablation) | CSL, PSC, and ACM · HRSC2016 and DOTA-v1.0 | AP75: CSL: 1.05 -> 77.76; 10.90 -> 38.71. PSC: 33.65 -> 79.20; 23.08 -> 39.35. ACM: 31.45 -> 86.33; 19.91 -> 41.44 | Helped |
| Add refine supervision (box loss) to ACM encoding supervision (ablation) | Dual-optimization with CenterNet · HRSC2016 and DOTA | AP75: 13.98 -> 88.33; 19.67 -> 42.83 | Helped |
| Add correct supervision (ACM loss) to box-loss supervision (ablation) | Dual-optimization with CenterNet · HRSC2016 and DOTA | AP75: 13.98 -> 88.33; 19.67 -> 42.83 | Helped |
Surprising findings
- Even state-of-the-art IoU-like methods (KFIoU, KLD) still exhibit severe angular mutation near the boundary, contrary to the long-held belief that joint-optimization solves boundary discontinuity.
- Under identical encoding, predicting the angle explicitly yields much weaker performance than predicting an implicit angular encoding (e.g., CSL explicit AP75 1.05 vs implicit 77.76 on HRSC2016 AP50/AP75 columns).
Limitations
- Experiments use only CenterNet and RoI-Transformer with a ResNet-50 backbone; generality to other detectors is not fully shown.
- Fair comparison to SOTA on DOTA is difficult due to differing settings, so the SOTA claim is only 'competitive'.
- Choice of omega=2 is justified by analysis and ablation partly deferred to supplementary material.
Experimental setup
Datasets: DOTA-v1.0 HRSC2016 UCAS-AOD
Benchmarks: DOTA-v1.0 HRSC2016 UCAS-AOD
Metrics: AP50 AP75 COCO-style AP
Baselines: CSL PSC GWD KLD KFIoU SkewIoU RoI-Transformer Gliding Vertex R3Det DCL
Builds on: CenterNet GWD KLD KFIoU PSC CSL
Closest papers by meaning
- Theoretically Achieving Continuous Representation of Oriented Bounding Boxes (CVPR 2024 · similarity 0.87)
- ReDiffDet: Rotation-equivariant Diffusion Model for Oriented Object Detection (CVPR 2025 · similarity 0.86)
- GauCho: Gaussian Distributions with Cholesky Decomposition for Oriented Object Detection (CVPR 2025 · similarity 0.85)
- FRED: Towards a Full Rotation-Equivariance in Aerial Image Object Detection (AAAI 2024 · similarity 0.84)
- ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects (CVPR 2025 · similarity 0.84)
- PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection (ICLR 2025 · similarity 0.84)
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.