BROWSE
Detail
Conference Paper
Toward open world object detection in various garbage environment datasets
Cited 0 time in 
Share 
Share
- Authors
- Hoseok Hwang, Aekyeung Moon
- Issue Date
- 2025-04
- Citation
- International Conference on Image and Video Processing (ICIVP) 2025, pp.1-5
- Language
- English
- Type
- Conference Paper
- Abstract
- To prevent illegal garbage dumping, we need to accurately detect garbage bags in various environments, which requires learning unknown classes. However, existing CNN-based methods cannot learn unknown classes, so they cannot quickly adapt to various environments. In this paper, we propose a semi-automatic OWOD annotation (SAOA) framework that automatically annotates unknown classes and updates the training dataset with human intervention, enabling rapid adaptation to various environments. Also, SAOA uses multi-class non-maximal suppression (M-NMS) to effectively suppress duplicate detection occurring in OWOD. Experimental results using the real Pay-As-You-Throw (PAYT) CCTV dataset demonstrate that SAOA detects unknown classes more effectively than CNN and suppresses duplicate detection effectively.
- KSP Keywords
- Open world, Rapid adaptation, Semi-Automatic, duplicate detection, human intervention, multi-class, non-maximal suppression, object detection, pay-as-you-throw