Stumbled onto a game changer for training my AI model's image recognition

I was stuck for days trying to get my model to tell apart different types of street signs, and it kept mixing up stop and yield signs. Finally decided to feed it only pictures taken in rain and fog, like 200 from a Portland dataset, and it suddenly clicked. The trick was making the training data intentionally bad or blurry, not perfect. Has anyone else tried this kind of adversarial training, or am I just lucky?

3 comments

3 Comments

kai_chen218d ago

So you're saying garbage data is actually good data now?

lewis.brian18d ago

Garbage in, garbage out still applies, but sometimes that garbage is actually just misunderstood treasure for training models.

carr.abby18d ago

Hold on, you guys are missing the real kicker here. It's not just about the rain and fog being bad data, its that the model is learning to ignore texture and color and focus on shape instead, which is way more reliable for signs. That Portland dataset probably has just the right amount of lens flare and water distortion to force it to see the actual outline of the stop sign versus the yield sign, not just red versus white.