Computer Vision News - September 2020

2 Oral Presentation 6 Rico Jonschkowski is a research scientist in robotics at Google Research. His work proposes an unsupervised method for estimating optical flow. He spoke to us ahead of his oral presentation. Optical flow is where you want to estimate the pixel-to-pixel correspondence between two images . Applied on video frames, it will give you the motion field in the video . Applied to stereo images, you get something called disparity , which directly translates into depth. It is a really useful thing that can be used for many different downstream tasks. Speaking of optical flow, some of our readers might want to read again our interview with Michael Black. This work is an approach to training unsupervised optical flow using raw videos without labeled data, which is important because it is very hard to gather labeled data for optical flow. There are almost no real datasets because on real video, you would need to know the models of everything that is moving around. Whereas if you can train from unlabeled video , you have access to a vast amount of data. What Matters in Unsupervised Optical Flow Best of ECCV 2020