This code is developed based on eyescream project with Torch: project site.
This code is the implementation of training and testing for S^2-GAN for the following paper:
Xiaolong Wang and Abhinav Gupta. Generative Image Modeling using Style and Structure Adversarial Networks. Proc. of European Conference on Computer Vision (ECCV), 2016. pdf
BibTeX:
@inproceedings{Wang_SSGAN2016,
Author = {Xiaolong Wang and Abhinav Gupta},
Title = {Generative Image Modeling using Style and Structure Adversarial Networks},
Booktitle = {ECCV},
Year = {2016},
}The trained models can be downloaded from dropbox.
The pre-processed dataset (NYUv2) including rgb images and TV-denoised Surface Normals in jpgs can be downloaded from dropbox.
The list of training files dropbox.
For training, one need to:
Update the path_dataset = '/scratch/xiaolonw/render_data/' in dataset.lua
Update the opt.save in train.lua for saving models For testing, one can download the models into the ssgan_models folder.
The code for Stucture-GAN is in structure-gan:
train.lua: training Stucture-GAN
test.lua: testing Stucture-GAN
ssgan_models/Structure_GAN.net is our trained modelThe code for Style-GAN without FCN constraints is in style-gan-nofcn:
train.lua: training Style-GAN
test.lua: testing Style-GAN (To run this you need to download the dataset)
ssgan_models/Style_GAN_nofcn.net is our trained modelThe code for Style-GAN with FCN constraints is in style-gan-fcn:
train_fcn.lua: training FCN for surface normal estimation
test_fcn.lua: testing FCN for surface normal estimation (To run this you need to download the dataset)
ssgan_models/FCN.net is our trained model
train_gan.lua: training Style-GAN
test_gan.lua: testing Style-GAN (To run this you need to download the dataset)
ssgan_models/joint_Style_GAN.net is our trained modelThe code for joint learning is in joint-ssgan:
train.lua: joint learning
test.lua: testing S^2-GAN
ssgan_models/joint_SSGAN.net is our trained model