Shen Sang, Manmohan Chandraker

This is the official code release of our ECCV2020 paper "Single-Shot Neural Relighting and SVBRDF Estimation". Please consider citing this paper if you find the code and data useful in your project.

1. PyTorch with CUDA support
2. Python3

We have included the pretrained models and some test cases inside this repo. Ensure that the folder structure under data is:

data
|-- models
|-- real
|-- output
|-- ...

Put your own test images under real. Then run test_real_env.py or test_real_pt.py to do inference. The estimated albedo, normal, roughness and depth, as well as the relighting images and videos will be shown under data/output.

Please download the synthetic dataset here. It contains all the materials and shape parameters (albedo, normal, roughness and depth) used for rendering. We also provide the script rendering.py for you to show how to render your own dataset. Unzip and rename it as Synthetic.

Make sure the structure is:

data
|-- datset
    |--Synthetic
        |--train
        |--test
|-- ...

Create the index file of all file names for the training set or test set by running python dataset/make_pkl.py.

1. Train the model for relighting under a single point light by running python train_pt.py.

2. Train the model for relighting under arbitrary environments and point light by running python train_env.py.

Evaluate the trained model by running python eval_env.py and python eval_pt.env.