Research map / Detection and segmentation / Open-set and few-shot detection
Progressive Exploration-Conformal Learning for Sparsely Annotated Object Detection in Aerial Images
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 PECL, a closed-loop framework that uses an RL-based conformal pseudo-label explorer and a multi-clue reward evaluator to adaptively select pseudo-labels for sparsely annotated aerial object detection.
Problem
Sparsely annotated object detection (SAOD) in aerial images, where fixed confidence thresholds for pseudo-label selection fail due to imbalanced predicted probabilities across object classes and sizes.
Approach
Formulates pseudo-label selection as a sequential decision problem: a policy network (conformal pseudo-label explorer) decides per candidate whether to accept it, using conformal non-conformity-based confidence, prototype similarity, and context; a Q-network (multi-clue selection evaluator) scores decisions with a reward based on entropy change and confidence-margin change, and both are co-optimiz
Key results
- On DOTA OBB at 1% label rate, ReDet w/ PECL reaches 63.72% mAP vs 48.10% for ReDet (+15.62), with similar gains for S2A-Net and OR-CNN.
- On DOTA at 5% label rate, PECL with ReDet reaches 67.06% mAP, surpassing Unbiased Teacher (64.74%), BRL, Co-mining (65.35%), and Region-based (65.71%).
- On HRSC2016 at 10% label rate, PECL improves OBB mAP to 80.65% (S2A-Net), 85.50% (OR-CNN), and 87.29% (ReDet).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PECL to ReDet | ReDet · DOTA OBB 1%, 2%, 5%, and 10% | mAP(%): 48.10 -> 63.72; 51.20 -> 66.58; 56.06 -> 67.06; 60.32 -> 69.36 | Helped |
| Add PECL to ReDet | ReDet · DOTA HBB 1%, 2%, 5%, and 10% | mAP(%): 48.57 -> 64.70; 52.69 -> 67.46; 57.93 -> 67.76; 61.22 -> 69.95 | Helped |
| Add PECL to ReDet | ReDet · HRSC2016 OBB and HBB at 5% and 10% | mAP(%): 58.28 -> 75.10; 69.36 -> 80.42; 79.66 -> 87.29; 83.36 -> 88.57 | Helped |
| Use ReDet w/ PECL instead of Region-based | DOTA OBB 5% sparse-annotated detection · DOTA OBB 5% | mAP(%): 65.71 -> 67.06 | Helped |
| Add the conformal pseudo-label explorer (ablation) | ReDet with sparse annotation · DOTA OBB 1% | mAP(%): 48.10 -> 60.84 | Helped |
| Add the multi-clue selection evaluator (ablation) | Explorer-only PECL variant · DOTA OBB 1% | mAP(%): 60.84 -> 61.31 | Helped |
| Add experience replay (ablation) | Explorer plus evaluator · DOTA OBB 1% | mAP(%): 61.31 -> 63.72 | Helped |
| Include confidence level in exploratory characteristics (ablation) | PECL conformal pseudo-label explorer · DOTA OBB 1% | mAP(%): 61.17 -> 63.72 | Helped |
| Include feature similarity in exploratory characteristics (ablation) | PECL conformal pseudo-label explorer · DOTA OBB 1% | mAP(%): 60.90 -> 63.72 | Helped |
| Include predicted probability in exploratory characteristics (ablation) | PECL conformal pseudo-label explorer · DOTA OBB 1% | mAP(%): 62.40 -> 63.72 | Helped |
| Use action space size 2 rather than size 1 (ablation) | PECL conformal pseudo-label explorer · DOTA OBB 1% | mAP(%): 60.81% vs 63.72% | Helped |
| Set prototype count to K=5 rather than K=10 (ablation) | PECL · DOTA OBB 1% | mAP(%): 63.72 -> 61.77 | Hurt |
Surprising findings
- Under 100% label rate on DOTA, PECL degrades performance (ReDet 76.25 vs 75.18), which the authors attribute to label noise introduced under full supervision.
- Adding CLIP to pseudo-label inference alone gives a large gain only modestly; PECL is needed for big improvements.
Limitations
- Designed for scenes with densely distributed objects; performance in general scenes may be suboptimal.
- Reinforcement learning increases overall training time.
- Localization precision is not sufficiently accurate according to visualization analysis.
Experimental setup
Datasets: DOTA HRSC2016
Benchmarks: DOTA HRSC2016
Metrics: mAP
Baselines: S2A-Net OR-CNN ReDet SOOD Unbiased Teacher Calibrated Teacher BRL Co-mining Region-based LSKNet-T
Builds on: ReDet OR-CNN S2A-Net Co-mining Unbiased Teacher
Closest papers by meaning
- ABBSPO: Adaptive Bounding Box Scaling and Symmetric Prior based Orientation Prediction for Detecting Aerial Image Objects (CVPR 2025 · similarity 0.85)
- Zero-Shot Aerial Object Detection with Visual Description Regularization (AAAI 2024 · similarity 0.85)
- As Pseudo-Label Free as Possible: Leveraging Adaptive Feature Generation for Sparsely Annotated Object Detection (AAAI 2025 · similarity 0.84)
- Plug and Play Active Learning for Object Detection (CVPR 2024 · similarity 0.84)
- FRED: Towards a Full Rotation-Equivariance in Aerial Image Object Detection (AAAI 2024 · similarity 0.84)
- PointSR: Self-Regularized Point Supervision for Drone-View Object Detection (CVPR 2025 · 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.