search

elliottwu / unsup3d

(CVPR'20 Oral) Unsupervised Learning of Probably Symmetric Deformable 3D Objects from Images in the Wild
MIT License
1.19k stars 192 forks source link

generalized to the clothing data set #14

Open fashionguy opened 4 years ago

fashionguy commented 4 years ago

Hi! I want to generalize this method to my cloth dataset generated by GAN. like this (mostly symmetry, about 1400) 102_AB 108_AB 110_AB

I try to train a model, but the outputs are very horibble. I am at a loss. Could you do me a favor, give me some suggestions. Thanks!

elliottwu commented 4 years ago

Hi! I suppose the depth maps are super noisy. Try adding a smoothing term and see if it helps to regularize the training. The smoothing term is not necessary in the face experiments, where the shape is easier to learn.

But I would be surprised if this model directly extends to clothes images. The biggest issue I imagine would be that most of the clothes images are frontal and in the symmetric viewpoint. This method requires a distribution of viewpoints, including non-symmetric viewpoints, ie, a collection of single-view images of clothes taken from various viewpoints.

fashionguy commented 4 years ago

Thanks again for your suggestions. I want to reconstruct the cloth using single image. Your suggestions let me a novice have a direction to try.

wangzhupi commented 2 years ago

hi ，could you tell me how to train on my own dataset？ thank you