More details about training

[Sanster]

Thanks for sharing the code and dataset. The encoder-only architecture makes DDCP faster and lighter than other methods, I really like the idea. I try to reimplement the paper, however, some training details are missing in the paper.

loss

1. What are α and β used for the pre-train model? In the utilsV4.py it's all equal 1

Experiments

2. What are the total epochs for the pre-train model? In train.py, the default epochs=300

[gwxie]

Hi,
1、please see here.

Document-Dewarping-with-Control-Points/Source/train.py

Line 116 in ba4639f

FlatImg.lambda_loss_a = 0.1

2、We have printed the EPOCH of pre-train model. see here

Document-Dewarping-with-Control-Points/Source/train.py

Line 93 in ba4639f

print("Loaded checkpoint '{}' (epoch {})"

[Sanster]

Thanks for your response!

Have you tried adding a semantic segmentation head? I tried to add an encoder to predict document mask, but the network does not converge.

[gwxie]

Thanks for your response!

Have you tried adding a semantic segmentation head? I tried to add an encoder to predict document mask, but the network does not converge.

Hi，
I've never done anything like this before.