ICCV Daily 2019 - Friday

something called curvature smoothing so that the segmentation will be more complete and smoother. How we do that? We present a novel mask pooling. It’s like max pooling, but you first multiply it by a mask and then you pool the maximum value. These two together form a new method. These are the core blocks.” Shir Gur is a PhD student at Tel Aviv University under the supervision of Professor Lior Wolf. He speaks to us ahead of his poster today. Shir tells us that they are trying to segment blood vessels to distinguish between background and vessels . The domain is a very sparse image. The photons you get are only from vessels or nothing, with low signal-to-noise ratio. They do it in an unsupervised way, so don’t need labels, annotations or annotator work which is very time consuming. How do they do it? Shir explains: “ The way we do it is actually taking something from the classical computer vision tasks which is active contour. We are using a classical method but using it in deep learning. Active contour is an iterative algorithm that tells you if a pixel should be segmented or not according to some criteria. That criteria is used as a loss function in the neural network. The morphological operators of the work that we are based on use 12 Poster Presentation DA I L Y Unsupervised Microvascular Image Segmentation Using an Active Contours Mimicking Neural Network "They do it in an unsupervised way, so don’t need labels, annotations or annotator work."