DeepInverseRendering

Source code of our paper:

Deep Inverse Rendering for High-resolution SVBRDF Estimation from an Arbitrary Number of Images

Duan Gao, Xiao Li, Yue Dong, Pieter Peers, Kun Xu, Xin Tong.

ACM Transactions on Graphics (Proceedings of SIGGRAPH 2019)*

More information (including our paper, supplementary, video and slides) can be found at My Personal Page.

If you have any questions about our paper, please feel free to contact me ([email protected])

Pretrained models

Our pretrained SVBRDF auto-encoder can be downloaded from here. Download the pretrained model and extract it into ./model/.

Dependencies

• Python (with opencv-python, numpy; test on Python 3.6)

• Tensorflow-gpu (test on tensorflow>1.10)

Usage

• Test mode (SVBRDF map as input)

python3 main.py 
   --N 20                                       # number of input images
   --checkpoint ../model/                       # pretrained model of auto-encoder
   --dataDir ../example_data/example_svbrdf     # folder contains a set of input images
   --logDir ../log_test_example                 # output folder
   --initDir  ../example_data/example_init      # folder contains initial SVBRDF maps or initial code 
   --network network_ae_fixBN                   # network architecture (default: network_ae_fixBN)
   
   --init_method svbrdf 
   --input_type svbrdf 
   --wlv_type load
   --wlvDir ../example_data/example_wlv

• Eval mode (captured images as input)

   python3 main.py                    
      --dataDir ../example_data/example_images    # LDR images are given in example_images    
      --input_type image 
      ... 

• SVBRDF format

  normal, diffuse, roughness, specular

  (diffuse map is in srgb color space (gamma 2.2), other maps are in linear color space)

• Command arguments:

-- input_type: ['image', 'svbrdf']
   
   'image':  a set of images, used in evaluate.
   
   'svbrdf':  a set of SVBRDFs, used in testing.
   
   
-- init_method: ['svbrdf', 'code', 'rand']

 'svbrdf': the estimated SVBRDF (using our encoder to embedding it into our latent space)
 
 'code':  the latent code (numpy array)
 
 'rand': random initialization the latent code


-- wlv_type: ['random', 'load']
   
   'random': random generate camera position and light position
   
   'load': load camera position and light position from file
   

Citation

If you use our code or pretrained models, please cite as following:

@article{gao2019deep,
  title={Deep inverse rendering for high-resolution SVBRDF estimation from an arbitrary number of images},
  author={GAO, DUAN and Li, Xiao and Dong, Yue and Peers, Pieter and Xu, Kun and Tong, Xin},
  journal={ACM Transactions on Graphics (TOG)},
  volume={38},
  number={4},
  pages={134},
  year={2019},
  month={July},
  publisher={ACM}
}