Deep Bilateral Learning for Real-Time Image Enhancements

Siggraph 2017

Visit our Project Page.

Michael Gharbi Jiawen Chen Jonathan T. Barron Samuel W. Hasinoff Fredo Durand

Maintained by Michael Gharbi (gharbi@mit.edu)

Tested on Python 2.7, Ubuntu 14.0, gcc-4.8.

Disclaimer

This is not an official Google product.

Setup

Dependencies

To install the Python dependencies, run:

cd hdrnet
pip install -r requirements.txt

Build

Our network requires a custom Tensorflow operator to “slice” in the bilateral grid. To build it, run:

cd hdrnet
make

To build the benchmarking code, run:

cd benchmark
make

Note that the benchmarking code requires a frozen and optimized model. Use hdrnet/bin/scripts/optimize_graph.sh and hdrnet/bin/freeze.py to produce these.

To build the Android demo, see dedicated section below.

Test

Run the test suite to make sure the BilateralSlice operator works correctly:

cd hdrnet
py.test test

Download pretrained models

We provide a set of pretrained models. One of these is included in the repo (see pretrained_models/local_laplacian_sample). To download the rest of them run:

cd pretrained_models
./download.py

Usage

To train a model, run the following command:

./hdrnet/bin/train.py <checkpoint_dir> <path/to_training_data/filelist.txt>

Look at sample_data/identity/ for a typical structure of the training data folder.

You can monitor the training process using Tensorboard:

tensorboard --logdir <checkpoint_dir>

To run a trained model on a novel image (or set of images), use:

./hdrnet/bin/run.py <checkpoint_dir> <path/to_eval_data> <output_dir>

To prepare a model for use on mobile, freeze the graph, and optimize the network:

./hdrnet/bin/freeze_graph.py <checkpoint_dir>
./hdrnet/bin/scripts/optimize_graph.sh <checkpoint_dir>

You will need to change the ${TF_BASE} environment variable in ./hdrnet/bin/scripts/optimize_graph.sh and compile the necessary tensorflow command line tools for this (automated in the script).

Android prototype

We will add it to this repo soon.

Known issues and limitations

• The BilateralSliceApply operation is GPU only at this point. We do not plan on releasing a CPU implementation.

• The provided pre-trained models were updated from an older version and might slightly differ from the models used for evaluation in the paper.

• The pre-trained HDR+ model expects as input a specially formatted 16-bit linear input. In summary, starting from Bayer RAW:

  1. Subtract black level.
  2. Apply white balance channel gains.
  3. Demosaic to RGB.
  4. Apply lens shading correction (aka vignetting correction).

  Our Android demo approximates this by undoing the RGB->YUV conversion and white balance, and tone mapping performed by the Qualcomm SOC. It results in slightly different colors than that on the test set. If you run our HDR+ model on an sRGB input, it may produce uncanny colors.