Computer Vision News - September 2022

42 Best Paper MIUA 2022 segmentation datasets. After performing different experiments and comparing their model against six state-of-the-art medical image segmentation methods, they demonstrated it outperformed them all – not only in one metric but several. What challenges did they face in the course of this work? “ The existing approaches perform very well already, so it’s tough to beat them! ” Javier declares. We note this is the first time we have heard someone say they were challenged to do better due to the success of the previous approaches! Vittorino responds: detect polyps even where the texture or color may not show them up. The team used the Pyramid Vision Transformer (PVTv2) , which utilizes the idea of pyramidal decomposition, a computer vision technique enabling the extraction of feature maps at different scales, suitable for dense prediction tasks like segmentation or object recognition. “ The selection of the backbone was very important, ” Vittorino tells us. “ PVTv2has anadvantageover transformers like ViT because it has a pyramidal structure that allows us to preserve some spatial information from the original image. ” The team reported results on five polyp