MhLiao / DB

A PyTorch implementation of "Real-time Scene Text Detection with Differentiable Binarization".
2.11k stars 481 forks source link

Do the text boundary boxes in annotation files have to be as close as possible to the text itself? #351

Closed btminh1820 closed 2 years ago

btminh1820 commented 2 years ago

I am transfer learning DB using my own custom dataset. However, the preannotations of the text instance boxes are not quite close to the text instance itself. I was wondering if it could affect the overall performance of DB?

btminh1820 commented 2 years ago

For example from my text boundary annotation: test

btminh1820 commented 2 years ago

@MhLiao

suhas004 commented 2 years ago

Hey @btminh1820 the bbox formation always takes in the post-processing step. You can go to structures -> representers -> seg_detector_representer.py and search for unclip_ratio and you can adjust the ratio there. The higher the ratio, the boxes will have more area comparatively

btminh1820 commented 2 years ago

@suhas004 Thanks for the greate advice!

dahalsweekar commented 1 year ago

@btminh1820 Please say how did you annotate your custom dataset?