search

gwxie / Dewarping-Document-Image-By-Displacement-Flow-Estimation

Dewarping Document Image By Displacement Flow Estimation with Fully Convolutional Network
MIT License
155 stars 35 forks source link

Folds and curves distortions not getting rectified #16

Open learning09 opened 1 year ago

learning09 commented 1 year ago 
          I tried running the code without using the given pretrained model, and trained from scratch for around 30 epochs. But the folds and curves are not removed from the image, only the background is cleaned as per below figure and below command :-

!python train.py --data_path_train=dataset/train/data1024_greyV2/color/ --data_path_validate=dataset/train/data1024_greyV2/color/ --data_path_test=dataset/shrink_1024_960/crop/ --parallel 0 --batch_size 2 --schema train --n_epoch 30 image

Could you please suggest the reason as to why the folds and curves are not getting removed and how to fix it?
Also, can i train on a small number of images instead of taking a huge dataset due to GPU limitations ?

Originally posted by @learning09 in https://github.com/gwxie/Dewarping-Document-Image-By-Displacement-Flow-Estimation/issues/2#issuecomment-1488860316