Design Automation of Out-of-Distribution Image Data Detectors
ABOUT THIS PROJECT
At a glance
Out-of-distribution (OOD) detection is an important topic in computer vision research. Despite the success of prior research in multi-class tasks, the performance of existing OOD detection methods on multi-label tasks is unsatisfactory. In our preliminary experiments, Machine Learning (ML)-based methods have shown much promise for multi-label classification tasks. However, the potential of ML-based methods has been downplayed in prior literature due to the lack of model selection, which is difficult because the OOD images are unknown. To resolve this challenge, we propose to generate OOD data with in-distribution data. Our overall research focus is on designing an automated model selection algorithm using generated OOD data that finds an optimal ML-based OOD detection model and corresponding hyperparameters to achieve robustness against OOD inputs, with a particular focus on autonomous driving applications.
Computer Vision, Out-of-Distribution Detection, Anomaly Detection