GradNet

A tensorflow 2.0 implementation of SIGGRAPH Asia 2019 paper: GradNet: Unsupervised Deep Screened Poisson Reconstruction for Gradient-Domain Rendering

The origin paper can be found GradNet: Unsupervised Deep Screened Poisson Reconstruction for Gradient-Domain Rendering

Thanks Mengtian Li and Jie Guo for their help.

Note that, this is no the official implementation.

Environment

• Python 3.7
• Tensorboard 2.0.0
• Tensorflow 2.0
• Openexr 1.3.2

How to train

Data prepartaion

• Generate images and the corresponding gradients by gradient domain render.You need to record normal, depth and albedo during the rendering procedure.
• Organize the directory structure like the description blow:

your_data_dir
  |--sence1
  |   |--grad
  |   |    |--sence1_1.gx
  |   |    |--sence1_1.gy
  |   |      ...
  |   |--sence1_1.exr
  |   |--sence1_1_feature.txt
  |   |  ...
  |
  |--sence2
   ...

• Modify thedata_dir to your train data directory , scene_list to your scenes in create_tfrecord_random.py.
• Run python create_tfrecord_random.py to generate tfrecords. It will randomly extract 15 patches of size 256x256 from each image and preprocess them according to the paper.

Train

• Copy some tfrecords to data/val/ as validation set.
• Modify the steps_per_epoch and val_steps in train.py.
• Run python train.py.
• You can run tensorboard --logdir log to watch the training process.

Test

• prepare test image, gradients and features.Put image and features into data/test/, put gradient image into data/test/grad/.
• Modify filepath_weights in test.py.
• Run python test.py.
• You can see the results in data/test/res/