There are a lot of fun things to do with un-annotated dataset, this is one of them.
I have taken the PASCAL dataset and thrown away all annotation. Next I wrote a few custom transformers, so that the pictures will be rotated, and the y values are rotations.
Add custom head, find optimum LR, freeze and unfreeze head and fit a few times.
After 5 minutes of training, I got some somewhat satisfying results.
That’s terrific! If you’re interested, it might be worth training that on a larger dataset (e.g. you could combine COCO + Pascal 2007 and 2012) and trying to get a model that works really well, then maybe create a little web app or client-side app that people can use to fix up their photos!
I can totally see this as python flask web app, deployed on pythonanywhere for 3 months for free.
For now, I’ll just write a quick blog post maybe, and move on to lecture 9mod7=2 — my plan is to make a small project/ for each lecture, and I’m already tad behind…