Research map / Detection and segmentation / Weakly supervised detection and segmentation
Relational Matching for Weakly Semi-Supervised 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.
Proposes a weakly semi-supervised oriented object detector that combines a few box-annotated images with many point-annotated images via teacher-student relational graph matching, rank distribution matching, and weakly supervised MIL/regression losses.
Problem
Oriented object detection needs costly box annotations, and performance drops when annotated data is scarce; point annotations on unannotated images are ambiguous for depicting oriented objects.
Approach
Builds relational graphs over proposals centered on annotated points, matches them between teacher and student under rotated augmented views weighted by modulated orientation difference, adds rank distribution matching on classification and regression, and applies multi-instance and regression-occupancy weak supervision to difficult points.
Key results
- On DOTA-v1.0 with 5% boxes, the method reaches 53.85 mAP vs 51.24 for Group R-CNN and 35.90 for the Faster R-CNN supervised baseline.
- On DIOR-R at 5% annotated data, mAP is 50.76 vs 48.33 for Group R-CNN.
- On DOTA-v1.0 at 20% annotated data, mAP is 64.48, which the authors say is comparable to other settings trained with 50% annotated data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full weakly semi-supervised method instead of Group R-CNN | Oriented object detection · DOTA-v1.0, partially annotated | mAP: 51.24 -> 53.85; 57.58 -> 59.69; 62.78 -> 64.48; 65.92 -> 67.04; 66.07 -> 68.02 | Helped |
| Use the full weakly semi-supervised method instead of Group R-CNN | Oriented object detection · DOTA-v1.5, partially annotated | mAP: 33.99 -> 46.29; 51.99 -> 53.27; 58.20 -> 59.10; 59.01 -> 60.17; 60.76 -> 61.92 | Mixed |
| Use the full weakly semi-supervised method instead of Group R-CNN or Dense Teacher | Oriented object detection · DIOR-R, partially annotated | mAP: 48.33 -> 50.76; 53.63 -> 54.69; 56.71 -> 57.53; 58.24 -> 58.80; 60.51 -> 61.72 | Helped |
| Add pseudo supervision on point-annotated images (ablation) | Proposed method · DOTA-v1.0, 5% annotated | mAP: 35.90 -> 43.91 | Helped |
| Add Rotation-Modulated Relational Graph Matching (ablation) | Proposed method with pseudo supervision · DOTA-v1.0, 5% annotated | mAP: 43.91 -> 48.78 | Helped |
| Add Relational Rank Distribution Matching (ablation) | Proposed method with graph matching · DOTA-v1.0, 5% annotated | mAP: 48.78 -> 50.22 | Helped |
| Add weakly supervised learning (ablation) | Proposed method with graph and rank matching · DOTA-v1.0, 5% annotated | mAP: 50.22 -> 53.85 | Helped |
| Add vertex matching (ablation) | Rotation-Modulated Relational Graph Matching · DOTA-v1.0, 5% annotated | mAP: 48.53 -> 50.51 | Helped |
| Add edge matching (ablation) | Rotation-Modulated Relational Graph Matching · DOTA-v1.0, 5% annotated | mAP: 48.53 -> 51.10 | Helped |
| Add both vertex and edge matching (ablation) | Rotation-Modulated Relational Graph Matching · DOTA-v1.0, 5% annotated | mAP: 48.53 -> 53.85 | Helped |
| Add classification rank matching (ablation) | Relational Rank Distribution Matching · DOTA-v1.0, 5% annotated | mAP: 49.58 -> 51.84 | Helped |
| Add regression rank matching (ablation) | Relational Rank Distribution Matching · DOTA-v1.0, 5% annotated | mAP: 49.58 -> 51.73 | Helped |
Limitations
- Relies on warm-up on annotated images and a Faster R-CNN base; gains on fully-annotated settings are smaller.
- Requires many hyperparameters (loss weights, thresholds tau_m) set by default values across experiments.
Experimental setup
Datasets: DOTA-v1.0 DOTA-v1.5 DIOR-R DOTA-v2.0
Benchmarks: DOTA-v1.0 DOTA-v1.5 DIOR-R
Metrics: mAP
Baselines: FCOS Faster R-CNN Dense Teacher Soft Teacher SOOD Group R-CNN
Builds on: Mean Teacher H2RBox SOOD Faster R-CNN Group R-CNN
Closest papers by meaning
- Point-Teaching: Weakly Semi-supervised Object Detection with Point Annotations (AAAI 2023 · similarity 0.91)
- Points As Queries: Weakly Semi-Supervised Object Detection by Points (CVPR 2021 · similarity 0.91)
- Partial Weakly-Supervised Oriented Object Detection (CVPR 2026 · similarity 0.90)
- Multi-clue Consistency Learning to Bridge Gaps Between General and Oriented Object in Semi-supervised Detection (AAAI 2025 · similarity 0.88)
- PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection (ICLR 2025 · 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)
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.