Computer Vision News - April 2021

5 Unsupervised Domain Adaptation from ... this can be easily related to the heart anatomy, oppositely to axial images that are oriented according to the thorax, making them subject to variation between individuals. However, these types of slices remain fundamental to investigate right ventricular quantification measurements for diseases that specifically involve right ventricular (RV) geometry, such as patients with congenital heart defects (CHD) e.g., tetralogy of fallot (TOF). These data (AX view) are uncommon and their corresponding annotations very scarce, making it a real challenge to train deep learning algorithms that can perform high quality tasks on them. Unfortunately, the problem is particularly complex because AX and SAX volume stacks describe a different physical volume, which means that each of the images contains information which is not present in the other image, and a rigid relative transformation between both images varies heavily frompatient to patient. The high variation - due to difference in resolution and field coverage - makes it impossible to simply get the transformation from each image and then just apply the average rotation to all AX images in order to close the domain gap. To ensure that this is really the case and further research is required to fill in the domain gap, the authors preliminarily try to directly apply a segmentation model trained on SAX images on axial slices (experiment 1). The experiment actually indicates a large distribution shift, similarly to an experiment (experiment 2) which shows that direct regression on the transformation parameters is also not sufficient. What does Julia Schnabel wish Computer Vision News for our 5th anniversary? Find out on page 31!