I built a new Optical Flow AI. Instead of being a cranky old man complaining about other people's mistakes, I decided to just fix things myself. Before Corona

I built a new Optical Flow AI. Instead of being a cranky old man complaining about other people's mistakes, I decided to just fix things myself. Before Corona hit, I had created a new data set and given a talk about all the stuff that current state of the art AIs for Optical Flow get wrong. I find the topic interesting, because those bugs directly lead to drones like Skydio crashing into thin wires or branches without leaves. Plus, I needed optical flow to work for a project, but the state of the art didn't work well enough. Then I thought "ah f* it" and so I built a new architecture that avoids the issues that I had been whining about. Now, 36 days of training later, it looks like I might have succeeded :) https://www.dropbox.com/s/3tfrs9sjm9heypk/example_selflow_hajo.jpg?dl=0 https://www.dropbox.com/s/3tfrs9sjm9heypk/example_selflow_ha... Top left and right are the two input images. Bottom center is the "ground truth", meaning the result that a perfect AI would produce. Bottom left is one of the State of the Art, according to the Sintel benchmark. Bottom right is my new architecture. You can see that the State of the Art tends to blur the two persons together and blur them with the background between them and the frame border. My approach uses monocular depth cues to correctly segment them into multiple movement layers.