Hand labeling training data vs letting the AI do it, big difference in my project

I used to spend whole weekends tagging images for my image recognition model, like 300 images at a time in my apartment in Seattle. Last month I tried using an auto-labeling tool with a confidence threshold set to 90%, then I only manually checked the ones it flagged. It cut my time down to about 2 hours, but I noticed the model got slightly worse at edge cases like blurry photos or weird lighting. Has anyone else seen a quality drop when switching to auto-labeling, or am I just being picky about the setup?

0 comments

0 Comments

No comments yet

Be the first to share your thoughts on this discussion.