Principle-centric AI with Adrien Gaidon - #575
The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)
00:00
*Their paper on teaching computer vision based on geometric formulas *
You can train it only from video. So you can just feed it raw vidios, and what it will learn if we learn to output point clouds or depth maps from there. And then, knowing the equations of geometry, we can recreate a past frame from the current frame by warping geometrically using an equation. Then, if the warping is correct, the pixtls a line. If it's not correct, they don't a line. The color are not the same. You're wrong because your neural netweights are wrong. We started from stereo set up with two cameras to just a singular camera to now all the crazy censor suetes.
Transcript
Play full episode
