Deep Reformulated Laplacian Tone Mapping

This is the implementation of Deep Reformulated Laplacian Tone Mapping.

Prerequisites for demo

• Pycharm 2019
• pretrained vgg16
• checkpoint
• demo tfrecord
• Tensorflow 1.9.0

Additional prerequisites for training and testing

• Laval Indoor dataset(EULA required)
• Luminance HDR

Instructions

Download this repo.

Network setup

1. Download pretrained vgg16.npy and place it under '/laplacianet/loss/pretrained/' folder.
2. Download the checkpoint(password: 9v3t if required). Unzip it and place all 4 files under '/laplacianet/checkpoint/demo/' folder.
3. Download the demo tfrecord(password: mcl0 if required). Unzip it and place it under '/laplacianet/dataset/tfrecord/' folder.
4. (optional) Download the HDR image in demo(password: frd0 if required). Unzip it and place it under '/laplacianet/dataset/demo/' folder.

If the above links are invalid, please check this link for checkpoint, tfrecord, and HDR images.

Pycharm setup

1. Download Pycharm. Go to File -> Open and choose the project where it's downloaded.
2. Go to File -> Settings. In the prompt window, select Project:laplacianet -> Project Interpreter on the left panel. At the top of the right panel, click the gear icon to add a Python Interpreter with environment Python 2.7.
3. In the virtual environment under the same panel, install the following dependencies:

• opencv-python. v 3.4.4.19
• tensorflow-gpu. v 1.9.0
• imageio. v 2.4.1 (need to install plugin to process hdr extension. Use the script imageio_download_bin freeimage in Pycharm terminal)
• easydict. v 1.9
• scipy. v 1.1.0
• matplotlib. v 2.2.3

Demo

In Pycharm, run /laplacianet/operation/test.py file.

Train

1. Contact the author to request full access to Laval Indoor dataset(~170GB).
2. Follow the data preprocessing steps specified on the paper to process the data.
3. Generate the label images. Luminance HDR and Photoshop are recommended.
4. Divide the data in train set and test set. Place the .hdr images oftrain set under '/laplacianet/dataset/train/hdr/' folder and the corresponding label images created from step 3 under /laplacianet/dataset/train/ldr/ folder. Place the .hdr images oftest set under '/laplacianet/dataset/test/hdr/' folder and the corresponding label images created from step 3 under /laplacianet/dataset/test/ldr/ folder.
5. To start training, in Pycharm, run /laplacianet/operation/train_high_layer.py to train the high frequency layer. run /laplacianet/operation/train_bottom_layer.py to train the low frequency layer. After the 2 layer's training accomplished, run /laplacianet/operation/train_all.py to fine tune the network. Modify the parameters on the top of the code to specify the layer level n. run /laplacianet/operation/tfboard.py file to monitor training using Tensorboard.

Test

1. In /laplacianet/operation/test.py file, modify the parameter mode to 'test'. Adjust the parameter level n as same as the training phase.
2. run /laplacianet/operation/test.py.