Research map / Detection and segmentation / Open-set and few-shot detection
OW-VAP: Visual Attribute Parsing for 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 OW-VAP, an LLM-free open world object detector built on YOLO-World that learns a visual attribute parser (MLP) with background pseudo-labels and probabilistic soft label assignment to detect unknown objects.
Problem
Open world object detection needs detectors to find unknown objects during inference; existing OVD-based methods depend on LLM-generated attribute descriptions whose accuracy and selection during incremental learning are problematic.
Approach
Replace LLM-generated attributes with coarse generic attribute sentences and train a lightweight MLP Visual Attribute Parser on known-matched and filtered background pseudo-labels, weighting pseudo-label loss with Weibull-based probabilistic soft labels (PSLA).
Key results
- On MOWODB Task 1, OW-VAP* reaches 86.6 U-Recall versus 73.5 for ovow* (deduplicated comparison), and OW-VAP reaches 58.8 U-Recall vs 41.5 for KTCN.
- In MOWODB Task 2, OW-VAP achieves 56.3 U-Recall, surpassing the zero-shot OVD upper bound (56.0) by 0.3.
- Task 1 U-AP of 13.0 on MOWODB, versus 10.4 for ovow*, with the zero-shot GT bound at 24.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add general attribute prompts and OOD probability (ablation) | Base Model (YOLO-World) · MOWODB Task 1 | U-Recall: 0.0 -> 45.6 | Helped |
| add general attribute prompts and OOD probability (ablation) | Base Model (YOLO-World) · MOWODB Task 1 | U-AP: 0.0 -> 10.4 | Helped |
| add Visual Attribute Parser (ablation) | general prompts + OOD probability · MOWODB Task 1 | U-Recall: 45.6 -> 58.5 | Helped |
| add Visual Attribute Parser (ablation) | general prompts + OOD probability · MOWODB Task 2 | U-AP: 7.5 -> 6.7 | Hurt |
| add Probabilistic Soft Label Assignment (ablation) | OW-VAP with Visual Attribute Parser · MOWODB Task 1 | U-AP: 10.5 -> 13.0 | Helped |
| add Probabilistic Soft Label Assignment (ablation) | OW-VAP with Visual Attribute Parser · MOWODB Task 2 | U-AP: 6.7 -> 9.3 | Helped |
| add Probabilistic Soft Label Assignment (ablation) | OW-VAP with Visual Attribute Parser · MOWODB Task 3 | U-AP: 6.2 -> 7.8 | Helped |
Surprising findings
- Visual attribute parser alone raises Task 1 U-Recall substantially but slightly lags in U-AP in later tasks, which PSLA then addresses.
- OW-VAP exceeds the zero-shot OVD generalization upper bound on U-Recall in Task 2 of MOWODB.
Limitations
- Performance on U-AP still trails the zero-shot OVD upper bound in Task 1 (13.0 vs 24.9).
- Relies on a fixed set of predefined generic attribute sentences and hyperparameters (alpha, delta) whose tuning is in the appendix.
- Benchmark results for baselines such as ovow and deduplicated comparisons depend on evaluation choices (OW-VAP* vs OW-VAP).
Experimental setup
Datasets: MOWODB SOWODB PASCAL VOC COCO
Benchmarks: MOWODB SOWODB
Metrics: U-Recall U-AP mAP WI A-OSE
Baselines: ORE OW-DETR ALLOW PROB CAT RandBox EO-OWOD Hyp-OW MEPU-FS SGROD
Builds on: YOLO-World OW-DETR MEPU-FS ORE
Closest papers by meaning
- OW-OVD: Unified Open World and Open Vocabulary Object Detection (CVPR 2025 · similarity 0.92)
- Detecting Open World Objects via Partial Attribute Assignment (CVPR 2025 · similarity 0.91)
- UMB: Understanding Model Behavior for Open-World Object Detection (NeurIPS 2024 · similarity 0.89)
- Semi-supervised Open-World Object Detection (AAAI 2024 · similarity 0.88)
- Hyp-OW: Exploiting Hierarchical Structure Learning with Hyperbolic Distance Enhances Open World Object Detection (AAAI 2024 · similarity 0.88)
- Open-World Objectness Modeling Unifies Novel Object Detection (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.