search

mcahny / Deep-Video-Inpainting

Official pytorch implementation for "Deep Video Inpainting" (CVPR 2019)
508 stars 95 forks source link

2 questions about the codes #27

Closed zosel260 closed 2 years ago

zosel260 commented 3 years ago

Dear authors, Thank you for sharing your codes. I'm very impressed from your results. While looking over your codes, there are 2 things I can't understand. Please look over these questions.

1) in demo_vi.py, line 167 if opt.double_size: prevmask = masks_[:,:,2]*0.5 => Why do you multiply 0.5 to mask for the double_size case?

2) in vinet.py, line 164

enc_input = torch.cat([masked_img, ones, ones*mask], dim=1)

For the encoder's inputs, Why do you insert ones channel ? I thinks ones channels has no information for inpainting. Why do you multiply ones channel to the mask? Is the ones * mask same with just the original mask?

I appreciate in advance.

mcahny commented 3 years ago

Hi, thanks for your interest. To answer your questions,

  1. The double_size case was trained with the mask where the hole region is filled with the value 0.5, and non-hole regions with 1.0. There is no special reason behind this choice.

  2. We follow the implementation of "Free-Form Image Inpainting With Gated Convolution - ICCV19". They spare the ones-channel for the additional user-sketch guidance, but this does not apply to our work, and thus would have minor effect with or without it.