BROWSE
Detail
Conference Paper
Leveraging Class-agnostic Object Proposals for Open-world Object Detection
Cited - time in 
Share 
Share
- Authors
- Assefa Seyoum Wahd, Minsu Jang, Seung-Ik Lee
- Issue Date
- 2023-08
- Citation
- International Conference on Robot and Human Interactive Communication (RO-MAN) 2023, pp.1-2
- Language
- English
- Type
- Conference Paper
- Abstract
- This paper proposes the use of class-agnostic foundation models, such as the segment anything model (SAM), for a realistic object detection scenarios where the model should detect bounding boxes for known and unknown objects. We utilize unknown bounding boxes proposed by SAM to train a faster R-CNN model-based OOD detector. Our method outperforms previous best approaches with a 36.9% improvement in FPR95 and an 8.57% increase in AUROC. However, our approach comes with a drop of 8.2% in known class mAP. Our findings highlight the potential of leveraging class-agnostic object proposal models for real-time OOD detection. Our proposed method adds OOD detection ability to the faster RCNN model without adding any computational overhead.
- KSP Keywords
- Bounding Box, CNN model, Detection ability, Faster r-cnn, Object Proposals, Object detection, Real-Time, model-based