fh2019ustc/DocGeoNet

Questions on OCR evaluation

Opened this issue · 2 comments

Soongja commented

Hi, I have a few questions on OCR evaluation.

  1. When evaluating OCR performance on DIR300 dataset(or DocUNet benchmark), the size of the predicted image and GT image are different. I suppose you have resized one of the two in advance. To which size did you resize the images?(predicted size or GT size?)

  2. Which tessdata(traineddata) did you use for Tesseract?(tessdata_fast or tessdata_best or tessdata)
    reference: https://tesseract-ocr.github.io/tessdoc/Data-Files.html

fh2019ustc commented

Hi, sorry for the late reply due to my health.
(1) I have uploaded the evalUnwarp.m in this repo.
(2) For the OCR evaluation, I do not resize the two images. Maybe you could explore the impact of resize operation.
(3) I didn't pay particular attention to this problem. I download the tesseract from the link and the version is 5.0.1.20220118.
Hope this helps~!

Soongja commented

Thank you for your reply! I was able to evaluate correctly.