ICCV Daily 2019 - Friday

new data, but also the essence of the segmentation that the network learns can produce very nice results between experiments. To sum up, Shir tells us what he is most proud of: “Taking a task that is usually supervised and making it unsupervised. This is a great achievement for me.” MICCAI 2019 DAILY 13 Shir Gur DA I L Y This is a jointworkwith theneuroscience department from Tel Aviv University. Shir says when they came to them the original data was 4D data. Like a movie of very, very sparse volumes of the brain. You could hardly see anything. In this work, they did the segmentation on the time collapse . They averaged the data (frames) over time. They don’t have labels for their task, but they can generate a large amount of it. He says this is the main reason they started this work. Shir tells us they have already published the next steps of this work in the Pacific Symposium on Biocomputing. It explores temporal segmentation of the vessels. You can see in the poster that images of the new domain look very sparse. Unlike previous methods that are supervised, shifting from one domain to another or from one experiment to another is very hard because the topology could change, and the experiment setting can change. You can see that because they were unsupervised, they can finetune their “Taking a task that is usually supervised and making it unsupervised. This is a great achievement for me.” If you want to learn more and ask Shir about his work, come along to his poster today at 15:30-18:00 in Hall B.