Computer Vision News - April 2021

7 is transformed back to the target domain. The authors first present a baseline transformation module (experiment 3) which is then expanded into an extended version (experiment 4) using two main techniques: the cycle consistency on forward and backward transformations, so that predicted transformations are invertible (SAX2AX - AX2SAX); the task-specific focus which guides the network in optimizing the transformation such that regions relevant to the segmentation task are included in the target grid. The final expanded network architecture, as employed in experiment 4, is described in the comprehensive figure below. What does Alyosha Efros wish Computer Vision News for our 5th anniversary? Find out on page 36! Unsupervised Domain Adaptation from ...