In this paper, the authors address the limitations in evaluating OOD detection methods by distinguishing between ID and OOD inputs.
They also propose using OOD unit tests, simple, synthetically generated image inputs, designed to assess OOD discovery weaknesses.
The authors analyze the performance of various architectures and Ood detection methods, revealing insights about model weaknesses and the impact of pretraining on OOD Detection performance.
A new paper was recently published in which the authors aim to address the limitation in evaluating oOD detection method.
By distinguishing betweenID and ODO inputs, OOD detecting methods enhance the model’s robustness and reliability in real-world applications.
Consequently, the evaluation of OOD detector methods is affected, resulting in underestimating the actual OOD detect performance and unjustly penalizing more effective OOD detectors.