We've been promised self-driving cars and autonomous drones, thanks to "breakthrough advancements" in AI. I just wanted to detect camera movement between two photos of my living room... Trying to debug why the "state of the art" AIs had failed me, I found beginners' mistakes, mathematical errors, crash bugs, numerical instability, wrong assumptions, overfitting, and an unhealthy fixation on GPU algorithms. I’ll show a new unreleased data set and I'm looking for collaborators to measure and document the issues with current deep learning AIs for Optical Flow.
Ort: Seminarraum INB (Geb. 64, 1.OG, Raum 17)