edwin-pan/MetaHDR

Model-Agnostic Meta-Learning for HDR Image Reconstruction. By learning the common structure between all LDR-to-HDR conversion tasks, our model is able to adapt it's predictions given extra exposures of a scene. This novel approach reframes LDR-to-HDR conversion as a meta-learning problem.

MetaHDR: Model-Agnostic Meta-Learning for HDR Image Reconstruction

Poster Video	Poster PDF

Directory Structure

.
├── configs                 # Stores yaml files for configuring training, evaluation, demo, etc.
├── data                    # Directory for data. See data.md for details on getting data.
├── materials               # Miscellaneous assets for the report, poster, video, etc.
├── scene_demo              # Small sample of input data used to demo adaptive forward-pass.
├── src                     
    ├── core                # Contains scripts used for configuration as well as the loss metrics used.
    ├── dataset             # Contains scripts used to read in the data, a data generator which will sample batches of data, as well as an hdr visualization script.
    └── model               # Contains scripts which define the UNet for the inner model, as well as functions for training, validating, and evaluating our model with MAML.
├── .gitignore
├── README.md
├── data.md
├── demo.py                 # Runs an adaptive forward pass to demonstrate deployed usage.
├── eval.py                 # Runs an evaluation of a trained model.
├── requirements.txt        # Required dependencies to run this model.
└── train.py                # Runs training iterations.

Author Contact:

Edwin Pan edwinpan@stanford.edu

Anthony Vento avento@stanford.edu

Getting Started

MetaHDR was implemented and tested on CentOS 7 with python >=3.9.0, and supports GPU and CPU computation. We have also provided a Colab demo for convenience.

First, clone the repository

git clone https://github.com/edwin-pan/MetaHDR.git

Next, install the requirements

pip install -r requirements.txt

Running the Demo

The demo code provided runs MetaHDR on any LDR image input. We have provided a sample scene input directory at ./scene_demo. LDR images need to be placed in ./scene_demo/LDR/ and the corresponding HDR labels should go to ./scene_demo/HDR/.

To run the demo,

python3 demo.py --input_folder ./scene_demo --output_folder ./scene_demo/output --cfg configs/config_demo.yaml

A new directory at ./scene_demo/output/ will contain the output HDR image.

Running the Evaluation

Make sure that data is downloaded and formatted correctly (see data.md).

To run evaluation,

python3 eval.py --model_dir <PATH-TO-TRAINING-OUTPUTS> --cfg <PATH-TO-CONFIG>

Running the Training

Make sure that data is downloaded and formatted correctly (see data.md). Training scripts will log the results in the output directory specified by the configuration file.

To run training,

python3 train.py --cfg <PATH-TO-CONFIG>

Google Colab

To quickly get a sense of how MetaHDR performs, we have provided a Google Colab notebook version of the demo that runs out of the box. The demo runs on the cloud for free.

Acknowledgement

This work was completed as a Final Project for EE 367 / CS 448I: Computational Imaging and Display at Stanford University. We would like to thank our professor, Dr. Gordon Wetzstein, for his valuable instruction throughout the quarter and our project mentor, Cindy Nguyen, for giving us insights into our project.

Citation

@inproceedings{metahdr2021,
  title={MetaHDR: Model-Agnostic Meta-Learning for HDR Image Reconstruction},
  author={Pan, Edwin and Vento, Anthony},
  month={March},
  year={2021}
}

References

A full list of references for this project can be found in the arxiv paper.