search

qcf-568 / DocTamper

[CVPR2023] Towards Robust Tampered Text Detection in Document Image: New Dataset and New Solution
127 stars 11 forks source link

No read the compressed image #36

Closed HHHuge closed 10 months ago

HHHuge commented 10 months ago

In line 76, the image is saved in JPEG format, and the JPEG information is read from the file (line 77). However, the RGB data is not re-read from the file.

image
qcf-568 commented 10 months ago

This might be a oversight in such script. As our model mainly rely on frequency domain clues to localize tamper, this has almost no affect on the performance and has no affect on the conclusion. The other pure RGB models in comparison may have a bit drop in performance with the extra compression.

HHHuge commented 10 months ago

Can you provide correct performance metrics?

qcf-568 commented 10 months ago

I plan to update it about two or three weeks later.

qcf-568 commented 9 months ago

After updating the final compression in both training and inference scripts, the metrics show an about 0.5 points fluctuation, which is within the normal range. So the original metrics are kept as valid ones. I'm very busy now and will upload the new model in April.