Do we need segmentation masks to train disCVAE?

[anno2020]

Hi Xinchen,

I just got to your wonderful paper Attribute2Image and I have a question about it. It will be great if you can answer it.

I saw that disCVAE decomposes the image generation into a foreground image generation and a background generation. The final image is the composition of the two. My question is do we need ground truth segmentation masks of training images to implement disCVAE?

It seem your code use the ground truth segmentation to train disCVAE?

eccv16_attr2img/scripts/train_lfw_disCVAE.lua

Line 222 in d7a4a64

local cur_mask = trainData[idx][2]:float():clone()

Thanks in advance!