Keywords: Out-of-Distribution detection, OOD detection, Vision transformer, anomaly detection, uncertain estimation
Abstract: Vision-based transformers have achieved comparable results to CNN models in tasks including object detection, image classification, and semantic segmentation. However, their performance in detecting Out-of-Distribution (OOD) samples during inference has not been fully evaluated. OOD detection plays an important role in safety-critical applications such as medical image analysis. In this paper, we evaluate 4 transformers on 2 open-sourced medical image datasets. Our results demonstrate the insufficient OOD detection performance of the transformers. Hence, future research in improving OOD detection should be encouraged.
Paper Type: validation/application paper
Primary Subject Area: Detection and Diagnosis
Secondary Subject Area: Uncertainty Estimation
Paper Status: original work, not submitted yet
Source Code Url: https://github.com/Shaunlipy/vision-ood
Data Set Url: https://github.com/Shaunlipy/vision-ood
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
TL;DR: We evaluate the OOD detection performance using vision-based transformers, and results reveal insufficient performance with promising research directions.