Computer Vision News - April 2021

6 Research The authors refer to their domain adaptation pipeline as an input adaptation type because it addresses the domain gap at the input level. It aims to alter the samples from the target domain (AX images) to bring the distribution of the target closer to that of the source (SAX images) and transform the result back into the target domain. The main DA approach is carried out by a Spatial Transformer which learns the 3D geometric relations between the two domains extracting a set of transformation parameters. These are used to transform the slice orientation to a short-axis view. Once this is completed, a task with pixel-wise predictions ( Segmentation Module ) can be performed on the SAX representation and the resulting image Hence, this study focuses on leveraging the knowledge and availability of SAX slices from two datasets to get optimal segmentation results on AX images of a TOF dataset with no ground truth labels. This is done by using an “unsupervised domain adaptation approach, which is able to align anisotropic image slice stacks with significantly different fields of view and small overlapping image regions”.