BROWSE
Detail
Conference Paper
Class-Wise Confidence Thresholding for OOD Detection in Robot Vision-based Applications
Cited - time in 
Share 
Share
- Authors
- Jihyun Hwang, Minsu Jang
- Issue Date
- 2024-06
- Citation
- International Conference on Ubiquitous Robots (UR) 2024, pp.1-2
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- This study proposes a novel method for detecting Out-Of-Distribution (OOD) data in image classification tasks, aimed at improving the trustworthiness of robot vision tasks e.g. surveillance or safety inspection. The proposed approach utilizes class-wise confidence thresholds, determined analytically through a grid search, to effectively identify data that falls outside the model’s training distribution. Experimental results demonstrate that the proposed method achieves competitive performance across various OOD detection scenarios, with significant improvements in Area Under the Receiver Operating Characteristic (AUROC) curve and False Positive Rate at 95% True Positive Rate (FPR95) compared to existing research. By accurately detecting model uncertainty, this study contributes to expanding the scope of indoor safety check robots, enhancing system reliability, safety, and efficiency. The proposed method’s feasibility in improving the trustworthiness of robot vision intelligence highlights its potential for real-world applications.
- KSP Keywords
- Competitive performance, Detecting model, False Positive(FP), False Positive Rate, Image Classification, Model uncertainty, Real-world applications, Receiver operating characteristic, Robot Vision, Safety check, Safety inspection