search

autodistill / autodistill-grounded-sam-2

Use Segment Anything 2, grounded with Florence-2, to auto-label data for use in training vision models.
https://docs.autodistill.com
Apache License 2.0
95 stars 13 forks source link

biased prompts #5

Open walternat1ve opened 3 months ago

walternat1ve commented 3 months ago

i noticed that you add a bias by prefixing each prompt with "a photo of ", this is not what the normal user expects. please remove it as this is influencing results. thx.

walternat1ve commented 3 months ago

furthermore i noticed that you use the class labels as prompts for florence-2, right?

another observation is that results differ a lot when prompts are combined. it seems to be better if one does prompt after prompt and then aggregate all target classes into one result.

danigarciaoca commented 1 month ago

Hi! I found that same problem. Check the issue I opened at Florence-2 repo; I think it really clears things up :smiley: