Research map / Detection and segmentation / Weakly supervised detection and segmentation
Point2RBox-v2: Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances
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-v2 trains a point-supervised oriented object detector using spatial layout among instances via Gaussian overlap loss, Voronoi watershed loss, plus edge loss, symmetry-aware consistency loss, and copy-paste augmentation.
Problem
Point-supervised oriented object detection (learning rotated boxes from only point annotations), where existing methods ignore spatial relationships among densely packed instances and often rely on SAM priors, one-shot examples, or two-stage pseudo-labeling.
Approach
Treats oriented objects as 2D Gaussians and penalizes their pairwise overlap (upper bound on size), and uses Voronoi tessellation of point annotations as markers for watershed to obtain per-instance regions that regress width/height (lower bound), combined with a rotation/flip/scale consistency loss between an image and its augmented view, an edge-snapping loss, and copy-paste augmentation.
Key results
- Point2RBox-v2 reaches 62.61% AP50 on DOTA-v1.0 in two-stage mode (FCOS on pseudo labels), outperforming P2RBox (59.04%) which uses a SAM prior.
- End-to-end Point2RBox-v2 achieves 51.00% AP50 on DOTA-v1.0, versus 34.07% for Point2RBox and 40.27% for Point2RBox+SK.
- Point2RBox-v2 achieves 86.15% AP50 on HRSC and 34.71% on FAIR1M in two-stage mode, and 64.00% on SKU110K end-to-end.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Point2RBox-v2 instead of Point2RBox | Point-supervised OOD, end-to-end training · DOTA-v1.0 | AP50: 34.07% vs. 51.00% | Helped |
| use Point2RBox-v2 instead of Point2RBox+SK | Point-supervised OOD, end-to-end training · DOTA-v1.0 | AP50: 40.27% vs. 51.00% | Helped |
| use Point2RBox-v2 instead of P2RBox | Point-supervised OOD, two-stage FCOS · DOTA-v1.0 | AP50: 59.04% vs. 62.61% | Helped |
| use Point2RBox-v2 instead of PointOBB-v2 | Point-supervised OOD, two-stage FCOS · DOTA-v1.5 and DOTA-v2.0 | AP50: 23.47%/18.15% higher | Helped |
| use Point2RBox-v2 instead of RBox-supervised FCOS | Point-supervised OOD, two-stage FCOS · HRSC | AP50: 86.15% vs. 88.99% | Hurt |
| use Point2RBox-v2 instead of PointSAM | Point-supervised OOD · DIOR | AP50: 44.45% vs. 46.20% | Hurt |
| use Point2RBox-v2 instead of H2RBox | Point-supervised OOD, two-stage FCOS · SKU110K | AP50: 65.64% vs. 57.15% | Helped |
| add Gaussian overlap loss and Voronoi watershed loss (ablation) | Point2RBox-v2 · DOTA-v1.0 | AP50: 0.00 -> 52.98 | Helped |
| add consistency loss after Gaussian and Voronoi losses (ablation) | Point2RBox-v2 · DOTA-v1.0 | AP50: 52.98 -> 54.26 | Helped |
| add edge loss after Gaussian, Voronoi, and consistency losses (ablation) | Point2RBox-v2 · DOTA-v1.0 | AP50: 54.26 -> 61.88 | Helped |
| add copy-paste augmentation (ablation) | Point2RBox-v2 · DOTA-v1.0 | AP50: +0.73% | Helped |
| set Gaussian overlap loss weight wO to 30 instead of 10 (ablation) | Point2RBox-v2 · DOTA-v1.0, end-to-end | AP50: 51.00 -> 45.88 | Hurt |
Surprising findings
- Point2RBox-v2 pseudo-labels train an FCOS detector to near RBox-supervised accuracy on six high-density DOTA categories (e.g., SH 86.9 vs 87.1).
Limitations
- Performance gap to RBox-supervised detection remains large for sparse categories such as bridge and soccer-ball field.
- Minimizing overlap as a soft constraint does not fully eliminate overlap between instances.
- Struggles on sparse objects not constrained by neighbors.
Experimental setup
Datasets: DOTA-v1.0 DOTA-v1.5 DOTA-v2.0 DIOR HRSC FAIR1M STAR SKU110K RSAR
Benchmarks: DOTA-v1.0 DOTA-v1.5 DOTA-v2.0 DIOR HRSC FAIR1M STAR SKU110K RSAR
Metrics: AP50
Baselines: Point2RBox Point2RBox+SK PointOBB PointOBB-v2 PointOBB-v3 P2RBox PointSAM H2RBox H2RBox-v2 Point2Mask
Builds on: Point2RBox H2RBox-v2 GWD PSC angle coder FCOS
Closest papers by meaning
- Point2RBox: Combine Knowledge from Synthetic Visual Patterns for End-to-end Oriented Object Detection with Single Point Supervision (CVPR 2024 · similarity 0.91)
- PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection (ICLR 2025 · similarity 0.88)
- WeakPCSOD: Overcoming the Bias of Box Annotations for Weakly Supervised Point Cloud Salient Object Detection (AAAI 2024 · similarity 0.87)
- PointOBB: Learning Oriented Object Detection via Single Point Supervision (CVPR 2024 · similarity 0.86)
- Relational Matching for Weakly Semi-Supervised Oriented Object Detection (CVPR 2024 · similarity 0.86)
- ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects (CVPR 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.