Open Set Recognition

Open-set recognition (OSR) tackles the challenge of classifying data points into known categories while reliably identifying samples from unknown classes, a crucial capability for real-world applications where exhaustive training data is unavailable. Current research focuses on improving the accuracy of both known-class classification and unknown-class detection, exploring techniques like hybrid models combining post-hoc OOD detection with test-time augmentation, and leveraging diverse feature learning and improved prototype generation. The development of robust OSR methods is vital for enhancing the safety and reliability of machine learning systems across various domains, from image recognition and face recognition to medical diagnosis and autonomous driving.