Trying to get a custom vision model to spot a specific part took me 3 weeks

I was building a system to check if a small metal bracket was installed right on an assembly line. The model kept getting confused by shadows and similar looking parts nearby. I thought it would take maybe 4 days to train and tune it, but it ended up needing over 3 weeks of messing with the training data and lighting setup. Some people say you should just throw more varied images at the problem, others say you need to control the environment perfectly first. Has anyone else hit a wall like this with a simple visual check?

4 comments

4 Comments

keith_bennett28d agoMost Upvoted

Man, I feel that. We had to add a cheap ring light to kill the shadows, which was way easier than collecting weeks of extra data. @emeryj66 is right about fixing the setup first, it just saves so much headache. Sometimes the simple hardware fix is the answer.

tessa_morgan28d ago

I read that controlling the environment first is usually cheaper than fixing it with data later.

emeryj6628d ago

Honestly that idea about controlling the environment first is so true. Tessa_morgan, I see it all the time with people trying to fix bad habits by just using willpower instead of changing their setup. It's way easier to just not buy the junk food than to try and resist it every day.

emeryj662d ago

Totally! I put the cookies on a high shelf in the pantry. Out of sight, out of mind, and I'm too lazy to get the step stool. Works every time.