Research map / Detection and segmentation / Weakly supervised detection and segmentation
Point2RBox: Combine Knowledge from Synthetic Visual Patterns for End-to-end Oriented Object Detection with 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.
Point2RBox trains an end-to-end, FPN-free oriented detector from single point annotations by combining knowledge from synthetic visual patterns placed around labeled points with transform self-supervision.
Problem
Oriented object detection normally needs costly rotated-box labels; point-supervised oriented detection lacks size and angle information, and existing point-to-HBox-to-RBox or point-to-mask-to-RBox pipelines are slow and less accurate.
Approach
Spread features from neighborhoods of each labeled point onto synthetic patterns with known boxes (recolored with point-derived colors) to supervise box regression, and enforce consistency of predicted RBoxes under flip, rotation and scale transforms of the input.
Key results
- Point2RBox-SK reaches 40.27% AP50 on DOTA-v1.0 with ResNet50 and 41.05% with CSPNeXt-l, versus 21.87% for the P2BNet+H2RBox-v2 pipeline.
- On DIOR, Point2RBox-SK achieves 27.34% AP50 (27.62% with CSPNeXt-l), above P2BNet+H2RBox-v2 at 23.61%.
- On HRSC, Point2RBox-SK reaches 79.40% AP50 (80.01% with CSPNeXt-l), slightly above KCR's 79.10% which uses RBox source data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use one sketch pattern per category instead of rectangles and circles with curve textures (ablation) | Point2RBox · DOTA-v1.0 | AP50: 34.07 -> 40.27 | Helped |
| Use one sketch pattern per category instead of rectangles and circles with curve textures (ablation) | Point2RBox · DIOR | AP50: 24.66 -> 27.34 | Helped |
| Use one sketch pattern per category instead of rectangles and circles with curve textures (ablation) | Point2RBox · HRSC | AP50: 78.77 -> 79.40 | Helped |
| Add curve textures to rectangles and circles (ablation) | Point2RBox · DOTA-v1.0 | AP50: 29.72 -> 34.07 | Helped |
| Add curve textures to rectangles and circles (ablation) | Point2RBox · DIOR | AP50: 22.23 -> 24.66 | Helped |
| Add curve textures to rectangles and circles (ablation) | Point2RBox · HRSC | AP50: 75.01 -> 78.77 | Helped |
| Add rotation and flip self-supervision (ablation) | Point2RBox-SK · DIOR | AP50: 21.34 -> 24.97 | Helped |
| Add rotation and flip self-supervision (ablation) | Point2RBox-SK · HRSC | AP50: 73.18 -> 78.13 | Helped |
| Add scale self-supervision (ablation) | Point2RBox-SK · DIOR | AP50: 24.97 -> 27.34 | Helped |
| Add scale self-supervision (ablation) | Point2RBox-SK · HRSC | AP50: 78.13 -> 79.40 | Helped |
| Add 10% point-coordinate noise (ablation) | Point2RBox-SK · DIOR | AP50: 27.34 -> 30.82 | Helped |
| Add 10% point-coordinate noise (ablation) | Point2RBox-SK · DOTA-v1.0 | AP50: 40.27 -> 39.60 | Hurt |
Surprising findings
- Adding annotation point noise of 10% improved AP50 on DIOR from 27.34% to 30.82%, which the authors attribute to reduced confusion with central circles in basketball courts.
Limitations
- Poor performance on categories with non-unique boundaries such as bridge, basketball court and storage/soccer field (BR/BC/SBF).
- Only applicable to FPN-free detectors like YOLOF-A1 because points carry no size information.
Experimental setup
Datasets: DOTA-v1.0 DIOR-RBox HRSC2016 synthetic visual patterns (SetRC, SetSK)
Benchmarks: DOTA-v1.0 DIOR-RBox HRSC
Metrics: AP50
Baselines: P2BNet+H2RBox-v2 P2BNet+H2RBox Point2Mask-RBox KCR YOLOF-A1 YOLOF-A5 H2RBox H2RBox-v2 RetinaNet S2A-Net
Builds on: YOLOF H2RBox-v2 P2BNet MMRotate
Closest papers by meaning
- Point2RBox-v2: Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances (CVPR 2025 · similarity 0.91)
- PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection (ICLR 2025 · similarity 0.90)
- PointOBB: Learning Oriented Object Detection via Single Point Supervision (CVPR 2024 · similarity 0.90)
- WeakPCSOD: Overcoming the Bias of Box Annotations for Weakly Supervised Point Cloud Salient Object Detection (AAAI 2024 · similarity 0.87)
- 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)
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.