Test-Time Adaptation for Segmentation
Single-image test-time adaptation improves segmentation robustness to data shift, reducing synthetic corruption errors by up to 25.12%.
Published Jun 14, 2023Paper ↗
Only vote on papers you've read. Sign in to vote.
Abstract
The thesis proposes novel single-image test-time adaptation methods for image segmentation based on a deep neural network. Test-time adaptation improves the robustness of the deep neural network to data shift using a single unlabelled image with no other data available. The methods are evaluated on two different deep segmentation networks, including a recent segmentation model trained on 1 billion segmentation masks, the SegmentAnything Model. The best approach reduces the segmentation error of the two models caused by synthetic corruptions by 15.47 % and 25.12 %.