Research map / Detection and segmentation / Weakly supervised detection and segmentation
PointOBB: Learning Oriented Object Detection via Single Point Supervision
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 PointOBB, the first method to generate oriented bounding boxes for aerial object detection from single point labels, using three views (original, resized, rot/flp) with scale-consistency and self-supervised angle learning.
Problem
Single point-supervised object detection has only produced horizontal boxes; oriented object detection in aerial images needs OBBs, and point supervision gives no scale or angle information.
Approach
Combine a MIL-based point-supervised detector with a resized view (scale-sensitive consistency loss) and a rotated/flipped view (self-supervised angle loss), plus a dense-to-sparse matching strategy that assigns dense angle predictions to sparse proposals, switched progressively during training.
Key results
- On DIOR-R, PointOBB with Oriented R-CNN reaches 38.08% mAP50 and 57.38% 8-mAP50, versus 22.59-23.61 for P2BNet+H2RBox(-v2) cascades (minimum improvement 14.47 points in mAP50).
- On DOTA-v1.0, PointOBB with Oriented R-CNN reaches 33.31% mAP50, surpassing the best plain two-stage point baseline (21.87%) by 11.44 points.
- Ablations show the SSC loss and DS matching each improve mIoU, and the gradual burn-in multi-view switching outperforms training the angle branch from the start (mAP50 38.08 vs 25.52).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SSC loss (ablation) | PointOBB · DIOR-R testing set | mAP50: 28.44 -> 32.98 | Helped |
| Add DS matching (ablation) | PointOBB · DIOR-R testing set | mAP50: 28.44 -> 30.63 | Helped |
| Add SSC loss and DS matching (ablation) | PointOBB · DOTA-v1.0 testing set | mAP50: 28.44 -> 33.31 | Helped |
| Apply MIL loss to the resized view (ablation) | PointOBB · DIOR-R testing set | mAP50: 34.94 -> 36.26 | Helped |
| Apply MIL loss to the rot/flp view (ablation) | PointOBB · DIOR-R testing set | mAP50: 34.94 -> 35.80 | Helped |
| Apply MIL loss to both enhanced views (ablation) | PointOBB · DIOR-R testing set | mAP50: 34.94 -> 38.08 | Helped |
| Attach angle-branch gradients to the other network parts (ablation) | PointOBB with gradual burn-in · DIOR-R testing set | mAP50: 25.52 -> 38.08 | Helped |
| Use gradual burn-in instead of introducing angle learning from the beginning (ablation) | PointOBB with gradient attachment · DIOR-R testing set | mAP50: 32.23 -> 38.08 | Helped |
| Group SSC scores by scale rather than proposal or ratio (ablation) | PointOBB · DIOR-R testing set | mAP50: 30.42; 35.78; 38.08 | Helped |
| Use a 10% point-label range rather than center-only labels (ablation) | PointOBB · DIOR-R testing set | mAP50: 32.11 -> 38.08 | Helped |
| Increase the point-label range from 10% to 20% (ablation) | PointOBB · DIOR-R testing set | mAP50: 38.08 -> 35.78 | Hurt |
Surprising findings
- Training the angle branch from the beginning (plain strategy) degrades MIL optimization, so the progressive burn-in schedule is necessary.
- Adding point noise (10% range) outperforms exact center points, since center points constrain the network's perception of object boundaries.
Limitations
- Suboptimal accuracy on ambiguous-boundary classes such as bridge, golf field, and overpass.
- Relies on the MIL paradigm, which focuses on discriminative parts rather than exact object boundaries.
- Experiments run on only two aerial datasets (DIOR-R and DOTA-v1.0).
Experimental setup
Datasets: DIOR-R DOTA-v1.0
Benchmarks: DIOR-R DOTA-v1.0
Metrics: mAP50 8-mAP50 7-mAP50 mIoU
Baselines: Rotated RetinaNet Rotated FCOS Oriented R-CNN WSODet H2RBox H2RBox-v2 Point2Mask-RBox P2BNet + H2RBox P2BNet + H2RBox-v2 Sun et al.
Builds on: P2BNet H2RBox-v2 Oriented R-CNN OICR MMRotate
Closest papers by meaning
- PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection (ICLR 2025 · similarity 0.94)
- Point2RBox: Combine Knowledge from Synthetic Visual Patterns for End-to-end Oriented Object Detection with Single Point Supervision (CVPR 2024 · similarity 0.90)
- PointSR: Self-Regularized Point Supervision for Drone-View Object Detection (CVPR 2025 · similarity 0.88)
- ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects (CVPR 2025 · similarity 0.88)
- Point2RBox-v2: Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances (CVPR 2025 · similarity 0.86)
- Mamba YOLO: A Simple Baseline for Object Detection with State Space Model (AAAI 2025 · 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.