Research map / Detection and segmentation / Weakly supervised detection and segmentation
BGHR: Bridging the Gap Between HBox-Supervised and RBox-Supervised Oriented Object Detection via Adaptive Fine-Grained Sample Mining
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 BGHR, an HBox-supervised oriented object detector that adds adaptive fine-grained sample mining (AFSM) using KLD-based prediction RBox assignment and an improved self-supervised loss on top of H2RBox.
Problem
HBox-supervised oriented object detection lags RBox-supervised methods; prior work closed the orientation-objective gap but ignored the gap in training sample selection, which picks noisy background samples.
Approach
AFSM pairs each GT HBox with its best-matching predicted RBox (chosen by KLD-based localization quality, optionally mixed with classification confidence via PRA) and selects positives from the intersection/union of GT-HBox and prediction-RBox regions, plus an improved symmetric-aware self-supervised loss Lss*.
Key results
- On DOTA-v1.0, BGHR reaches 71.44 AP50 vs 67.82 for H2RBox baseline (+3.62), and 77.65 AP50 with multi-scale testing.
- On DIOR-R, BGHR reaches 59.20 AP50 vs 57.00 for H2RBox, with AP 34.11.
- Ablation: AFSM and Lss* together raise DOTA AP75 from 33.67 to 38.05 with no added inference cost (same GFLOPs and FPS-level speed).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add AFSM and Lss∗ to the H2RBox baseline (ablation) | H2RBox · DOTA-v1.0 | AP75: 33.67 -> 38.05 | Helped |
| Add AFSM to the baseline Lws and Lss (ablation) | H2RBox · DOTA-v1.0 | AP75: 33.67 -> 36.52 | Helped |
| Replace Lss with symmetry-aware Lss∗ (ablation) | H2RBox · DOTA-v1.0 | AP75: 33.67 -> 35.64 | Helped |
| Use PM assignment instead of KO (ablation) | BGHR · DOTA-v1.0 | AP: 38.69 -> 39.75 | Helped |
| Use PM assignment instead of KO (ablation) | BGHR · DIOR-R | AP: 34.10 -> 34.61 | Helped |
| Set center sampling radius to 1.45 rather than removing center sampling (ablation) | BGHR · DOTA-v1.0 | AP: 24.19 -> 39.75 | Helped |
| Set center sampling radius to 1.45 instead of 1.40, 1.50, or 2.00 (ablation) | BGHR · DOTA-v1.0 | AP: 38.38; 39.04; 38.06 -> 39.75 | Mixed |
| Use BGHR instead of H2RBox | HBox-supervised oriented detection · DOTA-v1.0 | AP50: 67.82 -> 71.44 | Helped |
| Use BGHR instead of H2RBox | HBox-supervised oriented detection · DIOR-R | AP50: 57.00 -> 59.20 | Helped |
| Use BGHR instead of FCOS | oriented object detection · DOTA-v1.0 | AP50: 70.78 -> 71.44 | Helped |
| Use BGHR instead of FCOS | oriented object detection · DIOR-R | AP50: 58.60 -> 59.20 | Helped |
| Increase input image size to 1024 | BGHR · DIOR-R | AP50: 59.20 -> 60.70 | Helped |
Surprising findings
- BGHR (71.44 AP50 on DOTA) slightly exceeds the fully supervised RBox FCOS (70.78) despite using only HBox supervision.
Limitations
- Center-ness target still computed from GT HBox rather than fine-grained RBox.
- Hyperparameters (alpha, beta, radius) tuned by coarse search and sensitive to values (e.g., alpha=0.5, beta=2 drops AP to 24.83).
Experimental setup
Datasets: DOTA-v1.0 DIOR-R
Benchmarks: DOTA-v1.0 DIOR-R
Metrics: AP50 AP75 AP
Baselines: H2RBox EIE SAM BoxInst-RBox BoxLevelSet-RBox FCOS KLD RetinaNet GWD KFIoU PSC RepPoints
Builds on: H2RBox FCOS KLD (Yang et al. 2021d) ATSS
Closest papers by meaning
- ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects (CVPR 2025 · similarity 0.89)
- Point2RBox: Combine Knowledge from Synthetic Visual Patterns for End-to-end Oriented Object Detection with Single Point Supervision (CVPR 2024 · similarity 0.86)
- Point2RBox-v2: Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances (CVPR 2025 · similarity 0.84)
- Plug and Play Active Learning for Object Detection (CVPR 2024 · similarity 0.84)
- PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection (ICLR 2025 · similarity 0.84)
- Relational Matching for Weakly Semi-Supervised Oriented Object Detection (CVPR 2024 · similarity 0.83)
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.