Research map / Detection and segmentation / CNN features
Adaptive Important Region Selection with Reinforced Hierarchical Search for Dense 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 AIRS, a top-down hierarchical patch search over an FPN driven by an RL agent with evidential Q-learning (epistemic-uncertainty exploration) and a custom reward, which filters false-positive anchors for dense object detection.
Problem
One-stage dense object detectors produce many false positive anchors on difficult images with complex backgrounds because their non-adaptive criterion favors high recall.
Approach
An RL agent performs a coarse-to-fine hierarchical tree search over FPN patches, using an evidential Q-network whose epistemic uncertainty is added to the Q-value for exploration; the resulting binary RL masks filter candidate anchors before the detection head.
Key results
- AIRS reaches 48.3 AP on MS COCO vs 45.0 for GFocal and 47.8 for DINO-4scale; 35.6 AP on VOC 2012 vs 32.6 for GFocal.
- On the challenging COCO-style subset (APCH), AIRS scores 29.4 vs 25.4 for GFocal.
- AIRS trails two-stage detectors on large objects on MS COCO (APL 54.3 vs DINO 62.3), which the authors attribute to focusing on small objects.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add AIRS hierarchical region selection to GFocal | GFocal with ResNet-50 · MS COCO minival | AP: 45.0 -> 48.3 | Helped |
| Add AIRS hierarchical region selection to GFocal | GFocal with ResNet-50 · Pascal VOC 2012 | AP: 86.5 -> 88.7 | Helped |
| Add AIRS hierarchical region selection to GFocal | GFocal with ResNet-50 · Open Images V4 | AP: 45.8 -> 47.5 | Helped |
| Use IoU instead of DIoU + epistemic uncertainty as the positive-anchor criterion (ablation) | AIRS · MS COCO | AP: 47.6 -> 44.6 | Hurt |
| Use centerness instead of DIoU + epistemic uncertainty as the positive-anchor criterion (ablation) | AIRS · MS COCO | AP: 47.6 -> 42.4 | Hurt |
| Use GIoU instead of DIoU + epistemic uncertainty as the positive-anchor criterion (ablation) | AIRS · MS COCO | AP: 47.6 -> 44.3 | Hurt |
| Add epistemic uncertainty to GIoU-based selection (ablation) | AIRS · MS COCO | AP: 44.3 -> 46.7 | Helped |
| Add epistemic uncertainty to DIoU-based selection (ablation) | AIRS · MS COCO | AP: 45.4 -> 47.6 | Helped |
Surprising findings
- Removing epistemic uncertainty from the evidential Q-value causes a significant AP drop, which the authors tie to early termination of the top-down search.
- AIRS underperforms several baselines on large objects in MS COCO (APL), the trade-off for focusing on small and difficult objects.
Limitations
- Weaker on large objects in MS COCO compared with two-stage detectors.
- Theoretical bound relies on strong smoothness and concentration assumptions, with an action space of size D=5.
Experimental setup
Datasets: MS COCO PASCAL VOC 2012 Google Open Images V4
Benchmarks: MS COCO PASCAL VOC 2012 Open Images V4
Metrics: AP APS APM APL APCH mAP GFLOPS FPS
Baselines: Faster R-CNN Cascade R-CNN RepPoints TridentNet DETR Co-DETR EVA DINO RetinaNet FCOS ATSS SAPD
Builds on: GFocal RetinaNet Evidential Deep Learning Feature Pyramid Network DIoU
Closest papers by meaning
- Active Learning for Object Detection with Evidential Deep Learning and Hierarchical Uncertainty Aggregation (ICLR 2023 · similarity 0.85)
- Reinforcement Learning for Visual Object Detection (CVPR 2016 · similarity 0.85)
- Plug and Play Active Learning for Object Detection (CVPR 2024 · similarity 0.84)
- Towards Evidential and Class Separable Open Set Object Detection (AAAI 2024 · similarity 0.84)
- Portable Active Learning for Object Detection (CVPR 2026 · similarity 0.84)
- Expanding Low-Density Latent Regions for Open-Set Object Detection (CVPR 2022 · similarity 0.83)
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.