The way I was training AI image models was totally backwards

I spent like a month curating these huge datasets with hundreds of images each, thinking more data meant better results. Then a buddy who works at a tech shop in Portland told me he gets better outputs from just 15-20 really good, varied shots. He showed me his model for roof damage detection using only 18 photos and it picked out stuff mine missed. I tried it myself last week with a customer's house and the model actually recognized a cracked tile I hadn't even tagged. Turns out quality over quantity matters way more than people realize. Has anyone else noticed better results when they cut their training data way down?

1 comments

1 Comment

brooke7120d ago

That 18 photo example is wild, it really proves how much a tight, focused dataset beats throwing tons of random stuff at the model. I bet the key is making sure every single shot adds something unique instead of just repeating the same angle.