That said, I already have my covid project. All state of the art optical flow algorithms have problems seeing thin branches. You might have seen videos of Skydio drones crashing into branches, wires, or water. My goal is to fix that.
I started designing a new AI architecture in February and the main detection training finished mid-May. By now, I'm building the post-processing for guessing plausible fillings for occluded areas.
Plus, I'm now working from home, so my basement is already overheating from all those GPUs. I should probably pause AI for a while or I'll have to move the icebox to another floor.
1. You run statistics on a large text corpus to determine the most likely letter combinations that a font needs to display correctly.
2. You train a skip-word model on a large text corpus to evaluate if a sentence makes sense or not.
3. You train an AI to produce random sentences that match the word-pair-probabilities from #1 while being natural sentences according to #2.
=> You should end up with natural-sounding text with proper grammar that still contains all the important character pairs for font testing.