search

DengPingFan / FSGAN

47 stars 10 forks source link

The preprocessed dataset #1

Closed 649459021 closed 2 years ago

649459021 commented 2 years ago

I can't find the preprocessed dataset in google drive. The FS2K.zip in google drive folder seems to be unaligned?

ZhengPeng7 commented 2 years ago

Hi, as we mentioned in README, you can use the preprocessing codes provided here to obtain the exact aligned results.

We followed the APDrawingGAN to employ key point detection (MTCNN) and human segmentation (U-2-Net) sequentially. But codes in APDrawingGAN don't have an up-to-date preprocessing that's easy to use. So we strongly recommend you to directly use our preproc codes which are in the latest PyTorch version.

ZhengPeng7 commented 2 years ago

You can also directly download the FS2K_preproc.zip (preprocessing codes and the preprocessed data which have been split into train/test sets).

649459021 commented 2 years ago

You can also directly download the FS2K_preproc.zip (preprocessing codes and the preprocessed data which have been split into train/test sets).

Thank you very much for your patient reply. In addition, the input image size of model is resized from 250 or 318(original dataset size) to 512. The input image becomes a little blurry. Whether resize to 256 is more appropriate？

ZhengPeng7 commented 2 years ago

You are welcome. Actually, you can take a look at the Table. 1 in the original paper, the resolution is 299.74±95.07 × 273.56±38.67. Thus, though the resolution of more images in the dataset is closer to 256x256, we want to keep more information in those images of high resolution. You can also try resizing them to 256x256 in preproc codes and use them for training, I think the results will be similar to the existing ones.

649459021 commented 2 years ago

You are welcome. Actually, you can take a look at the Table. 1 in the original paper, the resolution is 299.74±95.07 × 273.56±38.67. Thus, though the resolution of more images in the dataset is closer to 256x256, we want to keep more information in those images of high resolution. You can also try resizing them to 256x256 in preproc codes and use them for training, I think the results will be similar to the existing ones.

Thanks so much again. I will try both sizes in experiment.