Computer Vision News - July 2022

55 Maria Papadomanolaki different time steps to refine and regress them [Figure 2]. In addition, information about the images’ edges is integrated during the training process, contributing to better image alignment and preservation of shapes. To handle properly the areas where there are no correspondences, a prior on the change regions is also added, guiding the model to relax the registration constraints in the areas of change. Our formulation can be integrated into any kind of fully convolutional architecture, providing at the same time fast inference performances . semantic segmentation on the available semantic categories that are presented in the different input dates, forming a multi- task framework. Concerning the image registration task, the developed method is a multi-step deformable registration scheme based on the expression power of deep fully convolutionalnetworks , regressingdirectly the spatial gradients of thedeformationand employing a 2D smooth transformer layer toefficientlywarpone image to theother, in an end- to-end fashion. The displacements are calculated in an iterative way, utilizing [Figure1] [Figure2]