MICCAI 2019 Wednesday

MICCAI 2019 DAILY 9 Ruibin Ma standard SLAM system. He explains: “The original name of a standard SLAM system is direct sparse odometry. It's purely CPU- based. It suffers the drift problem because of the aforementioned difficulties in colonoscopy images, so it won't work well. We combine it with a deep neural network which gives us some depth prediction of each camera frame, and we use that information in the real-time SLAM system. The SLAM system only does optimization or tracking, it doesn't have any prior knowledge of what the world should be, but a neural network does. This prior knowledge is like being given one frame and knowing what the scene should look like. We give that information from the neural network to the SLAM system.” The neural network also has a problem, which is that it is totally end to end. There is no optimization during the process. For example, if errors happen between two steps, they will be accumulated and there's no chance to fix them. That will cause problems when the sequence runs too long, so they use the real-time SLAM system to do optimization for the result from the deep neural network. Ruibin says they have created a win-win strategy. They have learnt from others but were the first to apply it to a recurrent neural network and a dense SLAM system. He concludes by telling us that colorectal cancer is the third most common cancer in men and the second in women worldwide. Colonoscopy is the golden standard for colon disease treatment at the first stage. To improve this, and specifically to decrease the rate of missing polyps, is crucial. To learn more about Ruibin’s work, please come along to his oral today at 14:00 and poster [W-9-M-305] at 15:30. "They resolve this by combining a deep neural network with a standard SLAM system"