Research map / Detection and segmentation / Weakly supervised detection and segmentation
ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects
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 ABBSPO, an HBox-supervised weakly supervised oriented object detector that adaptively scales GT horizontal boxes (ABBS) and adds a self-supervised symmetric prior angle loss (SPA), enabling training from both coarse and tight HBox annotations.
Problem
Weakly supervised oriented object detection from horizontal boxes degrades when GT HBoxes are tight (T-HBox) rather than coarse minimum-circumscribed boxes (C-HBox), and angle predictions are ambiguous under rotation/flip self-supervision.
Approach
Adaptively scale GT HBoxes per proposal according to predicted angle within a predefined scale range (ABBS) so predicted RBoxes can extend beyond tight HBoxes, and supervise angles with an SSIM-based symmetry loss on symmetric object categories (SPA loss) built on H2RBox-v2.
Key results
- On DIOR-R, ABBSPO reaches 64.33 3-AP50 and 59.70 AP50, surpassing H2RBox by 5.13 AP50 points and H2RBox-v2 by 3.03 AP50 points.
- On DOTA-v1.0 validation, ABBSPO achieves 69.26 AP50, outperforming H2RBox-v2 (67.69) and H2RBox (66.07).
- Ablation on DIOR-R shows ABBS raises 3-AP50 from 55.23 to 62.13 (with ABBS) and SPA adds further gains to 64.33 when combined.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ABBSPO instead of H2RBox-v2 | HBox-supervised oriented detection · DIOR-R test | 3-AP50: 55.23 -> 64.33 | Helped |
| Use ABBSPO instead of H2RBox-v2 | HBox-supervised oriented detection · DIOR-R test | AP50: 56.67 -> 59.70 | Helped |
| Use ABBSPO instead of H2RBox-v2 | HBox-supervised oriented detection · DOTA-v1.0 validation | 3-AP50: 63.47 -> 65.27 | Helped |
| Use ABBSPO instead of H2RBox-v2 | HBox-supervised oriented detection · DOTA-v1.0 validation | AP50: 67.69 -> 69.26 | Helped |
| Add ABBS to the method (ablation) | ABBSPO without ABBS or SPA · DIOR-R | 3-AP50: 55.23 -> 62.13 | Helped |
| Add SPA loss to the method (ablation) | ABBSPO without ABBS or SPA · DIOR-R | 3-AP50: 55.23 -> 58.77 | Helped |
| Add ABBS to the method (ablation) | ABBSPO without ABBS or SPA · DOTA-v1.0 | AP50: 67.69 -> 68.59 | Helped |
| Add SPA loss to the method (ablation) | ABBSPO without ABBS or SPA · DOTA-v1.0 | AP50: 67.69 -> 69.16 | Helped |
| Use Top-k proposal sampling only for SPA loss (ablation) | ABBSPO · DIOR-R | AP50: 58.93 -> 59.70 | Helped |
| Apply proposal sampling to losses other than SPA (ablation) | ABBSPO · DIOR-R | AP50: 59.70 -> 50.51 | Hurt |
| Apply proposal sampling to SPA and other losses (ablation) | ABBSPO · DIOR-R | AP50: 59.70 -> 50.91 | Hurt |
| Set ABBS scale range to 0.9–1.1 with interval 0.05 (ablation) | ABBSPO · DIOR-R | AP50: 59.70 -> 58.15 | Hurt |
Surprising findings
- Applying Top-k proposal sampling to the other losses (rot, flp, reg, cn, cls) significantly lowers performance, while sampling only for the SPA loss helps.
Limitations
- ABBS has limited effect on rectangular objects such as tennis courts and vehicles, and SPA gains are marginal for small objects with few pixels.
- Scale range must be tuned per HBox annotation type (1.0-1.5 for T-HBox vs near 1 for C-HBox).
- Orientation ambiguity remains for round objects; the paper relies on manually set horizontal orientations for some categories.
Experimental setup
Datasets: DIOR DIOR-R DOTA-v1.0 SIMD NWPU VHR-10
Benchmarks: DIOR-R DOTA-v1.0
Metrics: AP50 3-AP50
Baselines: H2RBox H2RBox-v2 PointOBB Point2RBox WSODet RetinaNet FCOS Oriented R-CNN GWD KLD KFIoU Oriented RepPoints
Builds on: H2RBox-v2 H2RBox FCOS PSC
Closest papers by meaning
- FRED: Towards a Full Rotation-Equivariance in Aerial Image Object Detection (AAAI 2024 · similarity 0.89)
- BGHR: Bridging the Gap Between HBox-Supervised and RBox-Supervised Oriented Object Detection via Adaptive Fine-Grained Sample Mining (AAAI 2025 · similarity 0.89)
- PointOBB: Learning Oriented Object Detection via Single Point Supervision (CVPR 2024 · similarity 0.88)
- WeakPCSOD: Overcoming the Bias of Box Annotations for Weakly Supervised Point Cloud Salient Object Detection (AAAI 2024 · similarity 0.87)
- Point2RBox: Combine Knowledge from Synthetic Visual Patterns for End-to-end Oriented Object Detection with Single Point Supervision (CVPR 2024 · similarity 0.87)
- Theoretically Achieving Continuous Representation of Oriented Bounding Boxes (CVPR 2024 · similarity 0.86)
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.