MIDL Vision 2021

out-of-distribution , ” Camila explains. “ That’s why in this paper we look at self-supervised models and how to use self-supervision loss calculated during testing to detect out-of- distribution samples . There are many different models and methods that can be used for out-of- distribution detection or uncertainty estimation. As well as having a method that generalizes to out-of- distribution data, you would use all these different methods because you need these extra quality checks . ” Previous research exists based on classification models, but Camila says it was a challenge to implement existing methods for semantic segmentation and make sure that they are working properly. There are two proxy tasks in the paper – contrastive learning and edge detection – chosen because they encourage the learning of geometric information that is relevant for the segmentation models. “ We weren’t sure of course when doing the experiments what the results would look like, ” Camila tells us. “ We had a strong assumption because it’s quite a basic expectation that the models will also fail at proxy tasks, but the most exciting part was when we got good results! ” There was one challenge that Camila was not expecting, but which could 11 Camila Gonzales VISION MIDL

RkJQdWJsaXNoZXIy NTc3NzU=