LANet: A Luminance Attentive Network with Scale Invariance for HDR Image Reconstruction

Overview

This is the author's reference implementation of the single-image HDR reconstruction using TensorFlow described in: "LANet: A Luminance Attentive Network with Scale Invariance for HDR Image Reconstruction"

The network architecture details are shown in "model.py" and the data processing is in "utils.py".

Prerequisites

• Python3
• numpy
• OpenCV >= 3.4
• Tensorflow == 1.13.1
• TensorLayer == 1.11.1

Usage

Pretrained model

The pretrained LANet checkpoints can be found in the checkpoints folder on Google Drive. The pretrained panoLANet checkpoints can be found in the checkpoints folder on Google Drive.

Inference

• Run your own images (using our trained LANet):

  cd LANet
  python ./src/main.py --phase test --gpu 0 --checkpoint_dir ./checkpoint_LANet/ --test_dir ./test/ --out_dir ./out/

• Run your own panoramas (using our trained panoLANet):

  cd panoLANet
  python ./src/main.py --phase test --gpu 0 --checkpoint_dir ./checkpoint_panoLANet/ --test_dir ./test/ --out_dir ./out/

Parameters and their description:

checkpoint_dir: path to the trained models.
test_dir: input images directory. This project provides a few sample images.
out_dir: path to output directory.

See main.py for more settable parameters.