Research map / Detection and segmentation / Open-set and few-shot detection
Debiased Novel Category Discovering and Localization
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 Debiased Region Mining (DRM), which fuses class-aware and class-agnostic (IoU/centerness-regression) RPN boxes, plus semi-supervised instance-level contrastive features and mini-batch K-means over-clustering, to discover and localize novel categories in COCO.
Problem
Object detectors trained on closed-set labels are biased toward seen classes, so they miss or misclassify unseen objects; the paper targets Novel Class Discovery and Localization (NCDL) without knowing the number of novel classes in advance.
Approach
Combine a class-aware RPN (accurate on known classes) with a class-agnostic RPN that replaces classification losses with IoU/centerness regression (better generalization to unknown objects), fuse their score-calibrated boxes via class-agnostic NMS, and cluster instance features from semi-supervised contrastive learning with mini-batch K-means over-clustering.
Key results
- On COCO train2014 with 1000 clusters, the method reaches AuC@0.5 0.1052 and discovers 57 classes versus 0.0971 and 45 for OLN.
- CorLoc reaches 0.9934 versus 0.9713 for OLN at both cluster settings.
- On COCO test2014 with 1000 clusters, DRM box plus contrastive features achieves 8.7 AP and 20.19 AP50, versus 1.9 AP and 5.2 AP50 for Faster R-CNN plus contrastive features.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DRM dual-RPN fused boxes instead of Faster R-CNN boxes (ablation) | object discovery with DINO features · COCO test2014 (#cluster=1000) | AP: 1.6 -> 3.4 | Helped |
| Use DRM dual-RPN fused boxes instead of OLN-Box proposals (ablation) | object discovery with DINO features · COCO test2014 (#cluster=1000) | AP: 1.5 -> 3.4 | Helped |
| Use contrastive features instead of DINO features (ablation) | Faster R-CNN box proposals · COCO test2014 (#cluster=1000) | AP: 1.6 -> 1.9 | Helped |
| Use contrastive features instead of DINO features (ablation) | OLN-Box proposals · COCO test2014 (#cluster=1000) | AP: 1.5 -> 1.5 | No clear effect |
| Combine DRM boxes with contrastive features (ablation) | object discovery · COCO test2014 (#cluster=1000) | AP: 3.4 -> 8.7 | Helped |
| Use 1000 clusters instead of 80 (ablation) | Ours (DRM) · COCO train2014 | AuC@0.2: 0.1203 -> 0.2381 | Helped |
| Use 1000 clusters instead of 80 (ablation) | Ours (DRM) · COCO train2014 | discovered classes: 20 -> 57 | Helped |
| Use Minibatch K-means instead of K-means (ablation) | clustering · COCO clustering, 10k samples and 80 clusters | clustering time: 16.371 -> 3.316 | Helped |
| Increase sample size (ablation) | clustering with Minibatch K-means and 80 clusters · COCO clustering | clustering time: 3.316 -> 3.997 | Hurt |
| Increase the number of clusters (ablation) | clustering with Minibatch K-means and 10k samples · COCO clustering | clustering time: 3.316 -> 17.282 | Hurt |
| Use over-clustering (ablation) | clustering · COCO clustering | AuC | Helped |
Surprising findings
- DualMem reproduced by the authors drops to AuC@0.5 0.0025 with 1000 clusters, far below its 80-cluster result.
- Replacing DINO features with the proposed contrastive features gives a large AP gain only when combined with DRM boxes (3.4 to 8.7 AP), while Faster R-CNN boxes gain little.
Limitations
- Results depend on a reproduced DualMem since its code was not released.
- K-means baseline does not scale, so clustering was run on a random 10,000-instance subsample of input instances.
- Evaluation uses only a VOC-to-COCO split, so generality to other known/unknown splits is untested.
Experimental setup
Datasets: Pascal VOC 2007 COCO 2014
Benchmarks: COCO 2014 train2014 (unknown classes) COCO test2014
Metrics: AuC@0.5 AuC@0.2 number of discovered classes CorLoc AP AP50
Baselines: Vplow DualMem OLN
Builds on: Faster R-CNN OLN DualMem DINO
Closest papers by meaning
- Learning to Discover and Detect Objects (NeurIPS 2022 · similarity 0.90)
- Generalized Category Discovery (CVPR 2022 · similarity 0.87)
- Few-Shot Object Detection via Association and DIscrimination (NeurIPS 2021 · similarity 0.87)
- Neighborhood Contrastive Learning for Novel Class Discovery (CVPR 2021 · similarity 0.86)
- ProxyDet: Synthesizing Proxy Novel Classes via Classwise Mixup for Open-Vocabulary Object Detection (AAAI 2024 · similarity 0.85)
- CAKE: Category Aware Knowledge Extraction for Open-Vocabulary Object Detection (AAAI 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.