Research map / Detection and segmentation / Open-set and few-shot detection
Hyp-OW: Exploiting Hierarchical Structure Learning with Hyperbolic Distance Enhances Open World 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 Hyp-OW, which learns a hierarchical known-class embedding in hyperbolic space with a SuperClass regularizer and uses hyperbolic centroid distances to relabel unmatched proposals as unknown in open world object detection.
Problem
Open world object detection struggles to identify unknown objects because there is no clear definition of "unknown" and models ignore semantic/hierarchical relations between known and unknown classes, leading to bias toward known labels and confusion with background.
Approach
Learn a hyperbolic embedding of known-class queries with a contrastive loss plus a SuperClass regularizer that groups classes by category, then relabel unmatched proposals as unknown when their hyperbolic distance to a known class centroid is below an adaptive threshold derived from matched known items in the batch.
Key results
- On the Hierarchical Split task 1, Hyp-OW reaches 34.9 U-Recall versus 29.4 for PROB.
- On the OW-DETR (low regime) split, Hyp-OW improves task 3 mAP to 46.2 versus 42.0 for PROB and task 4 mAP to 44.8 versus 39.9.
- On the Hierarchical Split, A-OSE for task 1 drops to 7,420 from 14,962 for PROB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Hyp-OW instead of PROB | Deformable DETR · Hierarchical Split Task 1 | U-Recall: 29.4 -> 34.9 | Helped |
| Use Hyp-OW instead of PROB | Deformable DETR · Hierarchical Split Task 2 | U-Recall: 43.9 -> 47.5 | Helped |
| Use Hyp-OW instead of PROB | Deformable DETR · Hierarchical Split Task 3 | mAP: 41.3 -> 44.3 | Helped |
| Use Hyp-OW instead of PROB | Deformable DETR · OW-DETR Split Task 1 | U-Recall: +6.3 | Helped |
| Use Hyp-OW instead of PROB | Deformable DETR · OWOD Split Task 1 | U-Recall: +4.1 | Helped |
| Replace hyperbolic distance with cosine distance (c=0) (ablation) | Hyp-OW · Hierarchical Split Task 1 | U-Recall: 34.9 -> 32.8 | Hurt |
| Replace hyperbolic distance with cosine distance (c=0) (ablation) | Hyp-OW · Hierarchical Split Task 1 | mAP: 49.9 -> 49.0 | Hurt |
| Vary curvature coefficient to c=0.2 or c=0.5 (ablation) | Hyp-OW · Hierarchical Split Task 1 | U-Recall: 34.9 -> 33.3; 32.3 | Hurt |
| Remove the SuperClass Regularizer (ablation) | Hyp-OW · Hierarchical Split Task 1 | U-Recall: 34.9 -> 32.0 | Hurt |
| Remove the SuperClass Regularizer (ablation) | Hyp-OW · Hierarchical Split Task 1 | mAP: 49.9 -> 50.0 | Mixed |
| Remove Adaptive Relabeling and classify all unmatched queries as unknown (ablation) | Hyp-OW · Hierarchical Split Task 1 | mAP: 49.9 -> 41.2 | Hurt |
| Remove Adaptive Relabeling and classify all unmatched queries as unknown (ablation) | Hyp-OW · Hierarchical Split Task 1 | U-Recall: 34.9 -> 34.7 | No clear effect |
Surprising findings
- Replacing hyperbolic distance with cosine distance (c=0) lowers both U-Recall and mAP, while larger curvature mainly degrades U-Recall.
- Relabeling all unmatched queries as unknown (PROB-style) keeps U-Recall nearly unchanged but substantially reduces known mAP (e.g., task 1 mAP 41.2 vs 49.9).
Limitations
- Hierarchical Split is designed by the authors, so gains are most pronounced on a dataset built to have strong hierarchy.
- Hyperparameters such as curvature and temperatures are tuned, and the low-regime U-Recall is on par with PROB except task 1.
Experimental setup
Datasets: PASCAL VOC 2007/2012 MS-COCO OWOD Split OW-DETR Split
Benchmarks: OWOD Split OW-DETR Split Hierarchical Split
Metrics: U-Recall mAP A-OSE
Baselines: ORE-EBUI OW-DETR PROB UC-OWOD OCPL 2B-OCD
Builds on: Deformable DETR OW-DETR PROB DINO Poincare embeddings
Closest papers by meaning
- OW-OVD: Unified Open World and Open Vocabulary Object Detection (CVPR 2025 · similarity 0.89)
- Open-World Objectness Modeling Unifies Novel Object Detection (CVPR 2025 · similarity 0.89)
- UMB: Understanding Model Behavior for Open-World Object Detection (NeurIPS 2024 · similarity 0.88)
- Semi-supervised Open-World Object Detection (AAAI 2024 · similarity 0.88)
- Exploring Orthogonality in Open World Object Detection (CVPR 2024 · similarity 0.88)
- Detecting Open World Objects via Partial Attribute Assignment (CVPR 2025 · similarity 0.88)
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.