Research map / Detection and segmentation / CNN features
Plug and Play Active Learning for 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 PPAL, a plug-and-play two-stage active learning method for object detection combining difficulty-calibrated uncertainty sampling and category-conditioned matching similarity diversity sampling, requiring no detector architecture or training changes.
Problem
Active learning for object detection is hard because uncertainty must jointly cover classification and localisation, and image similarity is ill-defined for multi-instance images; prior AL-for-detection methods modify detector architectures and training pipelines, limiting generalisation across dete
Approach
Two-stage AL: first re-weight per-object entropy by category-wise difficulty coefficients (combining classification and IoU-based localisation difficulty, tracked by EMA during training) to form a candidate pool; second, select a diverse batch with k-means++ seeded by k-center-greedy using CCMS, a category-conditioned instance-matching similarity between images.
Key results
- On VOC RetinaNet, PPAL reaches 72.6 mAP at 20% labelled data versus 66.6 for random sampling and 70.9 for entropy sampling.
- PPAL outperforms MIAL by 3.4 AP on COCO and 2.8 mAP on Pascal VOC in early AL rounds (<=3).
- PPAL beats random sampling on ResNet-101, MobileNetV2, and Swin-Tiny backbones across all labelled ratios on VOC RetinaNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DCUS instead of entropy sampling for candidate-pool selection (ablation) | RetinaNet with CCMS · Pascal VOC, 7.5% labeled images | mAP: 59.0±0.6 -> 60.8±0.5 | Helped |
| Use CCMS instead of global similarity for diversity sampling (ablation) | RetinaNet with DCUS · Pascal VOC, 7.5% labeled images | mAP: 58.8±0.6 -> 60.8±0.5 | Helped |
| Use CCMS instead of random selection for diversity sampling (ablation) | RetinaNet with DCUS · Pascal VOC, 7.5% labeled images | mAP: 60.3±0.7 -> 60.8±0.5 | Helped |
| Use difficulty-based category weights instead of class-frequency weights (ablation) | RetinaNet with CCMS · Pascal VOC, 7.5% labeled images | mAP: 59.3±0.6 -> 60.8±0.5 | Helped |
| Use CCMS instead of Jaccard similarity for diversity sampling (ablation) | RetinaNet with DCUS · Pascal VOC, 7.5% labeled images | mAP: 60.3±0.4 -> 60.8±0.5 | Helped |
| Use CCMS instead of FPN similarity for diversity sampling (ablation) | RetinaNet with DCUS · Pascal VOC, 7.5% labeled images | mAP: 59.4±0.3 -> 60.8±0.5 | Helped |
| Set candidate-pool expansion ratio δ to 4 (ablation) | PPAL on RetinaNet · Pascal VOC, 10% labeled images | mAP: δ=4: 66.2±0.4; δ=1: 64.4±0.4 | Helped |
| Use posterior probability instead of entropy as the uncertainty measure (ablation) | PPAL on RetinaNet · Pascal VOC, 7.5% labeled images | mAP: 60.8±0.5 -> 60.8±0.6 | No clear effect |
| Use probability margin instead of entropy as the uncertainty measure (ablation) | PPAL on RetinaNet · Pascal VOC, 7.5% labeled images | mAP: 60.8±0.5 -> 59.9±0.5 | Hurt |
| Use PPAL instead of random sampling | RetinaNet with ResNet-101 · Pascal VOC, 7.5% labeled images | mAP: 59.3±1.4 -> 64.3±0.5 | Helped |
| Use PPAL instead of random sampling | RetinaNet with MobileNetV2 · Pascal VOC, 7.5% labeled images | mAP: 31.2±2.8 -> 35.7±1.7 | Helped |
| Use PPAL instead of random sampling | RetinaNet with SwinTiny · Pascal VOC, 7.5% labeled images | mAP: 63.8±1.0 -> 68.4±0.3 | Helped |
Surprising findings
- Global feature-map similarity performs even worse than random selection for diversity sampling on multi-instance images.
- Class frequency-based difficulty weighting (DCUS-Freq) underperforms the detection-difficulty-based coefficient, as class AP does not strictly track training frequency.
Limitations
- SSD requires a different similarity (KL divergence on class probabilities) because features from different levels cannot be compared directly.
- Hyper-parameters (alpha, beta, delta, m0) were set empirically without careful tuning, and optimal settings vary across ablations.
Experimental setup
Datasets: COCO train2017 Pascal VOC 2007+2012
Benchmarks: COCO mini-val Pascal VOC test2007
Metrics: mAP AP (0.5:0.95)
Baselines: Random Entropy Core-set LearnLoss CDAL MIAL DivProto
Builds on: Core-set MIAL DDOD (difficulty-based detection loss) BADGE CDAL
Closest papers by meaning
- Portable Active Learning for Object Detection (CVPR 2026 · similarity 0.91)
- Uncertainty Meets Diversity: A Comprehensive Active Learning Framework for Indoor 3D Object Detection (CVPR 2025 · similarity 0.88)
- Active Domain Adaptation with False Negative Prediction for Object Detection (CVPR 2024 · similarity 0.87)
- Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty (NeurIPS 2026 · similarity 0.86)
- Language-conditioned Detection Transformer (CVPR 2024 · similarity 0.85)
- Open-Det: An Efficient Learning Framework for Open-Ended Detection (ICML 2025 · 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.