search

switchablenorms / DeepFashion_Try_On

Official code for "Towards Photo-Realistic Virtual Try-On by Adaptively Generatingā†”Preserving Image Content"ļ¼ŒCVPRā€˜20 https://arxiv.org/abs/2003.05863
825 stars 252 forks source link

Question about SSIM evaluaton, training epoch and the input of content fusion module #14

Closed LeoXing1996 closed 4 years ago

LeoXing1996 commented 4 years ago

Thanks for this great work, however, I still have some questions about this work.

  1. How should I evaluate SSIM? I directly calculate SSIM score on image reconstruction task (reference and target are from the same image). However, the pretrained model and my own trained model (20 epoch) get 0.7980 and 0.7594 on test set.

  2. Does this model only need 20 epochs training? In default options, this model would be trained for 200 epochs. I found that SSIM score is still increasing after 20 epochs and reached 0.7653 in epoch 40.

  3. In released training and inference code, average skin color (skin_color) of each class area is used in the input of content fusion module instead of synthesized clothing mask (M_c^S) mentioned in paper. šŸ¤” https://github.com/switchablenorms/DeepFashion_Try_On/blob/a628ca5612431ab5e8c0e729265e9d3453d71575/ACGPN_train/models/pix2pixHD_model.py#L323

Amazingren commented 3 years ago

Hi @LeoXing1996, I am also want to replicate the SSIM metric for evaluation. However, I have no idea about how to do this by calculating SSIM between generated Images and ground truth images. So how should I find the ground truth images firstly? It would be great help if you can share some advice with me on this issue~

LeoXing1996 commented 3 years ago

Hi @LeoXing1996, I am also want to replicate the SSIM metric for evaluation. However, I have no idea about how to do this by calculating SSIM between generated Images and ground truth images. So how should I find the ground truth images firstly? It would be great help if you can share some advice with me on this issue~

In SSIM evaluation, the model performs a reconstruction task, like the training process. Therefore, GT images are from the test set.