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.

Build

Make sure to use Android-NDK 12b. Version 13 and above cause known issues with Tensorflow for Android. We tested on Android SDK 25.0.1.

Change the path to tensorflow, android-sdk and android-ndk in WORKSPACE to match your environment.

Change the target_features and generators_args parameters in android/BUILD to match your device/emulator.

Then build:

bazel build //android:hdrnet_demo

Plug your smartphone in, or launch an emulator, then install the package on device:

bazel mobile-install //android:hdrnet_demo

To add a new operator, freeze and optimize a TF model:

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

Then add a folder in android/assets containing the optimized_graph.pb file. Also add the *.bin weights for the guidemap.

Finally, add the name of this folder to the filters_array in android/res/values/arrays.xml.

The rendering shader that implements the slice and apply operation is in android/assets/camera_preview.frag.

• Camera2 API yields YUV images, we convert to RGB to feed the TF input in android/jni/convert_image_hl.cxx. The processing could be sped up by training a model on YUV images directly.

• OpenGL float textures are in the range [0,1] so we normalize the affine weights. (see android/jni/convert_output_hl.cxx). The shader needs to undo this normalization.

Known issues and limitations

• The BilateralSliceApply operation is GPU only at this point. We do not plan to release 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 HDR+ pretrained model has a different input format (16-bits linear, custom YUV). It will produce uncanny colors if run on standard RGB images.