Research map / Detection and segmentation / Weakly supervised detection and segmentation
PointSR: Self-Regularized Point Supervision for Drone-View 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 PointSR, a point-supervised drone-view object detector that uses temporal-ensembling anchor adaptation and informative negative sampling to produce higher-quality pseudo boxes from point labels.
Problem
Generating accurate pseudo-box labels from point-level supervision for densely packed small objects in drone-view images, where rigid (static) positive sampling and random negative sampling make MIL-based pseudo-box training unstable and inefficient.
Approach
Self-regularized sampling: a Temporal-Ensembling Encoder keeps a per-class aspect-ratio prototype (EMA over pseudo-box shapes) to dynamically adjust anchors, and an Informative-Sample Collector selects high-loss negatives after NMS to refine pseudo boxes within a WSDDN-style MIL classifier.
Key results
- PointSR reaches average AP50 of 42.5% on DroneVehicle, 38.7% on UAVDT and 35.0% on VisDrone with point supervision, beating prior best by 7.2, 5.4 and 2.6 points.
- Ablation on DroneVehicle: TE Encoder alone +0.92 AP50, Informative Sampling alone +5.26 AP50 over Point-to-Box baseline, combined 42.53 AP50.
- Robustness to point perturbation: AP50 variance 0.171 on DroneVehicle vs 0.935 for Point-to-Box.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PointSR instead of PointOBBv2 | Faster R-CNN trained with generated pseudo boxes · VisDrone (Point-Supervision) | AP50: 32.4 -> 35.0 | Helped |
| Use PointSR instead of Point-to-Box | Faster R-CNN trained with generated pseudo boxes · DroneVehicle (Point-Supervision) | AP50: 35.3 -> 42.5 | Helped |
| Use PointSR instead of Point-to-Box | Faster R-CNN trained with generated pseudo boxes · UAVDT (Point-Supervision) | AP50: 33.3 -> 38.7 | Helped |
| Use PointSR instead of Point-to-Box | Faster R-CNN trained with generated pseudo boxes · DroneVehicle with point-label perturbations | Variance of AP50: 0.935 -> 0.171 | Helped |
| Add TE Encoder (ablation) | Point-to-Box baseline · DroneVehicle (Point-Supervision) | AP50: 35.26 -> 36.18 | Helped |
| Add TE Encoder (ablation) | Point-to-Box baseline · DroneVehicle (Point-Supervision) | mIoU: 86.0 -> 86.7 | Helped |
| Add Informative Sampling (ablation) | Point-to-Box baseline · DroneVehicle (Point-Supervision) | AP50: 35.26 -> 40.87 | Helped |
| Add Informative Sampling (ablation) | Point-to-Box baseline · DroneVehicle (Point-Supervision) | mIoU: 86.0 -> 88.1 | Helped |
| Add TE Encoder and Informative Sampling (ablation) | Point-to-Box baseline · DroneVehicle (Point-Supervision) | AP50: 35.26 -> 42.53 | Helped |
| Add TE Encoder and Informative Sampling (ablation) | Point-to-Box baseline · DroneVehicle (Point-Supervision) | mIoU: 86.0 -> 88.9 | Helped |
| Set TE Encoder λ to 0.5 rather than 0.1 (ablation) | PointSR · DroneVehicle | AP50: 34.2 -> 42.5 | Helped |
| Set TE Encoder λ to 0.5 rather than 0.3 (ablation) | PointSR · DroneVehicle | AP50: 40.5 -> 42.5 | Helped |
Surprising findings
- Point-supervised detector trained on pseudo boxes outperforms box-supervised Faster R-CNN in some cases per qualitative figure.
- Sampling ratio of informative negatives significantly affects results, with sensitivity to lambda (best 0.5, 34.2 to 42.5 AP50 range).
Limitations
- Relies on simulated point annotations (Rectified Gaussian perturbation) rather than real human labels on all datasets.
- Gains on VisDrone are smaller (2.6 to 3.3 AP50) and mIoU on some classes remains low.
Experimental setup
Datasets: DroneVehicle VisDrone UAVDT
Benchmarks: DroneVehicle VisDrone UAVDT
Metrics: AP50 mIoU variance of AP50
Baselines: PLUG Point2Rbox PointOBB PointOBBv2 Point-to-Box Faster R-CNN (box-supervised upper bound)
Builds on: WSDDN Point-to-Box Faster R-CNN focal loss
Closest papers by meaning
- PointOBB: Learning Oriented Object Detection via Single Point Supervision (CVPR 2024 · 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)
- Leveraging Imagery Data with Spatial Point Prior for Weakly Semi-supervised 3D Object Detection (AAAI 2024 · similarity 0.86)
- RemDet: Rethinking Efficient Model Design for UAV Object Detection (AAAI 2025 · similarity 0.84)
- BEV-MAE: Bird’s Eye View Masked Autoencoders for Point Cloud Pre-training in Autonomous Driving Scenarios (AAAI 2024 · similarity 0.84)
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.