Research map / Detection and segmentation / Weakly supervised detection and segmentation
PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point 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 PointOBB-v2, a single-branch point-supervised oriented detector that learns a class probability map from non-uniform positive/negative assignment, then uses weighted PCA and vector-constrained suppression to produce pseudo rotated boxes from points.
Problem
Generating accurate oriented bounding box pseudo-labels from single point annotations for oriented object detection in aerial images, where existing modular methods like PointOBB are slow, memory-hungry, and fail in dense scenes.
Approach
Replace PointOBB's teacher-student structure with a single network that outputs a class probability map trained with distance-based positive/negative sample assignment; estimate orientation via weighted PCA over probability-weighted grid points, and separate neighboring objects with a vector constraint.
Key results
- On DOTA-v1.0 with ReDet detector, PointOBB-v2 reaches 44.85 mAP50 vs 33.95 for PointOBB (+10.90), and 41.68 vs 30.08 with Rotated FCOS.
- Pseudo-label generation takes 1.43 hours vs 22.28 hours for PointOBB (15.58x faster).
- On DOTA-v1.5 and DOTA-v2.0 (ReDet), PointOBB-v2 achieves 36.39 and 27.22 mAP50 versus 11.24 and 6.03 for PointOBB; memory around 8GB without RoI limits.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace PointOBB with PointOBB-v2 using ReDet | ReDet · DOTA-v1.0 | mAP50: 33.95 -> 44.85 | Helped |
| Replace PointOBB with PointOBB-v2 using ReDet | ReDet · DOTA-v1.5 | mAP50: 11.24 -> 36.39 | Helped |
| Replace PointOBB with PointOBB-v2 using ReDet | ReDet · DOTA-v2.0 | mAP50: 6.03 -> 27.22 | Helped |
| Replace PointOBB with PointOBB-v2 pseudo-label generation | PointOBB · Pseudo-label generation | training time (hours): 22.28 -> 1.43 | Helped |
| Add negative sample assignment (ablation) | PointOBB-v2 label assignment · DOTA-v1.0 | mAP50: 23.62 -> 44.75 | Helped |
| Add middle-region negative assignment (ablation) | PointOBB-v2 label assignment · DOTA-v1.0 | mAP50: 44.75 -> 44.85 | Helped |
| Use weighted rather than probabilistic PCA sampling (ablation) | PointOBB-v2 · DOTA-v1.0 | mAP50: 41.40 -> 44.85 | Helped |
| Add the vector constraint module (ablation) | PointOBB-v2 · DOTA-v1.0 | mAP50: 27.88 -> 44.85 | Helped |
| Set PCA sampling size to 7 rather than 5 (ablation) | PointOBB-v2 · DOTA-v1.0 | mAP50: 43.35 -> 44.85 | Helped |
| Increase PCA sampling size from 7 to 9 (ablation) | PointOBB-v2 · DOTA-v1.0 | mAP50: 44.85 -> 44.17 | Hurt |
| Increase PCA sampling size from 7 to 11 (ablation) | PointOBB-v2 · DOTA-v1.0 | mAP50: 44.85 -> 43.63 | Hurt |
| Add 10% point annotation noise (ablation) | PointOBB-v2 · DOTA-v1.0 | mAP50: 44.85 -> 42.30 | Hurt |
Surprising findings
- Negative sample assignment is far more important than positives alone: positive-only training yields 23.62 mAP50 versus 44.75 with negatives added.
- Middle-region negatives alone (without other negatives) drop performance to 18.21 mAP50.
Limitations
- Requires at least two point annotations per image to define negatives via nearest-neighbor distance, which may degrade performance on extremely sparse scenes.
- Several hyperparameters (e.g., assignment radii) are dataset-specific and may need adjustment elsewhere.
- Performance degrades with point annotation noise (mAP50 drops about 2.27 average at 10% shift).
Experimental setup
Datasets: DOTA-v1.0 DOTA-v1.5 DOTA-v2.0
Benchmarks: DOTA-v1.0 DOTA-v1.5 DOTA-v2.0
Metrics: mAP50 mIoU training hours memory (GB)
Baselines: PointOBB Point2RBox-RC Point2RBox-SK P2BNet+H2RBox P2BNet+H2RBox-v2 Point2Mask-RBox
Builds on: PointOBB Rotated FCOS Oriented R-CNN ReDet MMRotate
Closest papers by meaning
- PointOBB: Learning Oriented Object Detection via Single Point Supervision (CVPR 2024 · 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)
- Point2RBox-v2: Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances (CVPR 2025 · similarity 0.88)
- Relational Matching for Weakly Semi-Supervised Oriented Object Detection (CVPR 2024 · similarity 0.87)
- PointSR: Self-Regularized Point Supervision for Drone-View Object Detection (CVPR 2025 · similarity 0.87)
- Semi-supervised 3D Object Detection with PatchTeacher and PillarMix (AAAI 2024 · similarity 0.85)
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.