Research map / Detection and segmentation / Open-set and few-shot detection
OW-OVD: Unified Open World and Open Vocabulary 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 OW-OVD, a YOLO-World-based detector that unifies open-vocabulary and open-world object detection via visual-similarity attribute selection (VSAS) and hybrid attribute-uncertainty fusion (HAUF) for unknown object prediction.
Problem
Existing detectors handle either open-vocabulary detection (zero-shot, but cannot actively discover unknowns) or open-world detection (unknown discovery with incremental learning, but no zero-shot capability); their combination is largely unexplored.
Approach
Starts from a standard OVD detector and selects LLM-generated attributes whose similarity distributions are similar between annotated and unannotated regions (via JSD plus a diversity constraint), then predicts unknowns in parallel with the OVD classifier by fusing attribute similarity with known-class uncertainty, leaving standard OVD inference unchanged.
Key results
- On M-OWODB, OW-OVD doubles unknown recall over ORTH (+25.4 U-Recall in Task 1 per paper text); on S-OWODB Task 1 it gains +15.3 U-Recall and +4.3 mAP over prior SOTA.
- U-mAP on S-OWODB Task 1 reaches 23.0 vs 7.5 for MEPU-FS (+15.5).
- On VOC 10+10 incremental setting, OW-OVD reaches 87.2 mAP vs 72.3 for ORTH (+14.9).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OW-OVD instead of KTCN | M-OWODB Task 1 · M-OWODB Task 1 | U-Recall: 41.5 -> 50.0 | Helped |
| Use OW-OVD instead of SKDF | S-OWODB Task 1 · S-OWODB Task 1 | U-Recall: 60.9 -> 76.2 | Helped |
| Use OW-OVD instead of MEPU-FS | S-OWODB Task 1 · S-OWODB Task 1 | U-mAP: 7.5 -> 23.0 | Helped |
| Use OW-OVD instead of ORTH | PASCAL VOC incremental learning · PASCAL VOC 10+10 | mAP: 72.3 -> 87.2 | Helped |
| Add out-of-distribution probability (ablation) | OW-OVD with all attributes · M-OWODB | U-mAP: 4.3 -> 5.9 | Helped |
| Add visual similarity attribute selection (ablation) | OW-OVD with out-of-distribution probability · M-OWODB | U-mAP: 5.9 -> 6.9 | Helped |
| Add attribute similarity restriction (ablation) | OW-OVD with visual similarity attribute selection · M-OWODB | U-mAP: 6.9 -> 7.1 | Helped |
| Add known-class uncertainty (ablation) | OW-OVD with attribute similarity restriction · M-OWODB | U-mAP: 7.1 -> 7.4 | Helped |
| Weight the top attribute scores (ablation) | OW-OVD with known-class uncertainty · M-OWODB | U-mAP: 7.4 -> 8.6 | Helped |
| Use all generated attributes instead of the selected top attributes (ablation) | OW-OVD · M-OWODB | U-mAP: 8.6 -> 4.3 | Hurt |
Surprising findings
- Using all LLM-generated attributes for unknown prediction gives very low U-mAP (4.3 on M-OWODB), far below the upper bound, showing attribute selection is necessary.
- Incremental performance is nearly insensitive to the split (10+10 vs 19+1 gap of 0.4 AP), unlike baselines such as PROB (gap of 6.1).
Limitations
- Attributes for unknown objects must be selected from descriptions of known categories, since unknown names are unavailable.
- Unknown-object U-mAP remains modest on M-OWODB (8.6 in Task 1).
Experimental setup
Datasets: MS COCO PASCAL VOC M-OWODB S-OWODB
Benchmarks: M-OWODB S-OWODB PASCAL VOC incremental splits (10+10, 15+5, 19+1)
Metrics: U-Recall U-mAP mAP
Baselines: ORE OW-DETR PROB CAT RandBox ORTH Hyp-OW MEPU-FS SGROD SKDF KTCN ILOD
Builds on: YOLO-World FOMO OW-DETR CLIP TAL
Closest papers by meaning
- OW-VAP: Visual Attribute Parsing for Open World Object Detection (ICML 2025 · similarity 0.92)
- UMB: Understanding Model Behavior for Open-World Object Detection (NeurIPS 2024 · similarity 0.92)
- Detecting Open World Objects via Partial Attribute Assignment (CVPR 2025 · similarity 0.91)
- Open-World Objectness Modeling Unifies Novel Object Detection (CVPR 2025 · similarity 0.91)
- How to Evaluate the Generalization of Detection? A Benchmark for Comprehensive Open-Vocabulary Detection (AAAI 2024 · similarity 0.90)
- Exploring Orthogonality in Open World Object Detection (CVPR 2024 · similarity 0.90)
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.