Research map / Detection and segmentation / Open-set and few-shot detection
Looking Beyond the Known: Towards a Data Discovery Guided 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 CROWD, which reframes open-world object detection as combinatorial data discovery (submodular conditional gain mining of unknown RoIs) plus submodular-loss representation learning to reduce known/unknown confusion and forgetting.
Problem
Open-world object detection suffers from semantic confusion between known and unknown classes and catastrophic forgetting of previously known classes, yielding low unknown recall and degraded known-class accuracy.
Approach
Model unknown discovery as a submodular conditional gain (SCG) targeted selection over RoI proposals (CROWD-Discover), and fine-tune with a family of submodular-derived losses (CROWD-Learn) that separate known from unknown embeddings while keeping known-class clusters compact.
Key results
- On M-OWODB, CROWD reaches 38.5 mAP in Task 4 vs 37.9 for OrthogonalDet, and up to 69.6 U-Recall in Task 3 vs 29.1 for OrthogonalDet.
- On S-OWODB, CROWD achieves 46.4 mAP in Task 4 vs 46.2 for OrthogonalDet and 50.5 vs 24.6 U-Recall in Task 1 relative to OrthogonalDet.
- On PASCAL VOC IOD 10+10, CROWD reaches 73.5 mAP vs 69.41 for OrthogonalDet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CROWD’s discovery and learning framework | OrthogonalDet on M-OWOD · M-OWOD Task 1 | U-Recall: 24.6 -> 57.9 | Helped |
| Add CROWD’s discovery and learning framework | OrthogonalDet on S-OWOD · S-OWOD Task 1 | U-Recall: 24.6 -> 50.5 | Helped |
| Add CROWD’s discovery and learning framework | OrthogonalDet on M-OWOD · M-OWOD Task 4 | mAP Both: 37.9 -> 38.5 | Helped |
| Add CROWD’s discovery and learning framework | OrthogonalDet on S-OWOD · S-OWOD Task 4 | mAP Both: 46.2 -> 46.4 | Helped |
| Add CROWD-L during IOD fine-tuning | PASCAL VOC 10 + 10 IOD · PASCAL VOC 10 + 10 | mAP: +5.9% | Helped |
| Use Graph-Cut conditional gain for discovery only (ablation) | OrthogonalDet on M-OWOD · M-OWOD Task 1 | U-Recall: 24.6 -> 57.0 | Helped |
| Use Facility-Location conditional gain for discovery only (ablation) | OrthogonalDet on M-OWOD · M-OWOD Task 1 | U-Recall: 24.6 -> 50.7 | Helped |
| Use Log-Determinant conditional gain for discovery only (ablation) | OrthogonalDet on M-OWOD · M-OWOD Task 1 | U-Recall: 24.6 -> 56.4 | Helped |
| Use Facility-Location conditional gain for learning only (ablation) | OrthogonalDet on M-OWOD · M-OWOD Task 1 | U-Recall: 24.6 -> 25.0 | No clear effect |
| Use Graph-Cut conditional gain for learning only (ablation) | OrthogonalDet on M-OWOD · M-OWOD Task 1 | U-Recall: 24.6 -> 24.3 | Hurt |
| Use Log-Determinant conditional gain for learning only (ablation) | OrthogonalDet on M-OWOD · M-OWOD Task 1 | U-Recall: 24.6 -> 22.7 | Hurt |
| Combine CROWD-D and CROWD-L (ablation) | OrthogonalDet on M-OWOD · M-OWOD Task 1 | U-Recall: 24.6 -> $57.9_{\pm0.33}$ | Helped |
Surprising findings
- Using a Graph-Cut-based SCG selection for unknown mining with the same budget gave the largest U-Recall gains, while the best CROWD-L variant was Facility-Location based rather than Graph-Cut.
Limitations
- CROWD-D still injects a small set of spurious exemplars into the selected unknown pool.
- Reliance on SCG as the combinatorial formulation may be restrictive.
Experimental setup
Datasets: M-OWODB S-OWODB MS-COCO PASCAL-VOC
Benchmarks: M-OWODB S-OWODB PASCAL VOC incremental object detection (10+10, 15+5, 19+1)
Metrics: U-Recall mAP Wilderness Impact A-OSE
Baselines: ORE OST OW-DETR UC-OWOD ALLOW PROB CAT RandBox OrthogonalDet
Builds on: OrthogonalDet Faster R-CNN ORE Submodular Mutual Information / Total Information losses
Closest papers by meaning
- Open-World Objectness Modeling Unifies Novel Object Detection (CVPR 2025 · similarity 0.89)
- OW-OVD: Unified Open World and Open Vocabulary Object Detection (CVPR 2025 · similarity 0.89)
- Detecting Open World Objects via Partial Attribute Assignment (CVPR 2025 · similarity 0.89)
- Hyp-OW: Exploiting Hierarchical Structure Learning with Hyperbolic Distance Enhances Open World Object Detection (AAAI 2024 · similarity 0.87)
- Semi-supervised Open-World Object Detection (AAAI 2024 · similarity 0.86)
- UMB: Understanding Model Behavior for Open-World Object Detection (NeurIPS 2024 · similarity 0.86)
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.