DiffHDRsyn
DiffHDRsyn copied to clipboard
An end-to-end learning framework for HDR reconstruction using multi-exposure images
DiffHDRsyn
This repository contains the python implementation for
full paper "End-to-End Differentiable Learning to HDR Image Synthesis for Multi-exposure Images" at AAAI 2021
short paper "Differentiable HDR Image Synthesis using Multi-exposure Images" at DiffCVGP NeurIPSW 2020
If you find our paper or code useful, please cite our papers.
Requirements
The dataset can be downloaded from here
Pretrained weights can be downloaded from here
The code was tested under the following setting:
- pytorch >= 1.2.0
- torchvision >= 0.4.0
- scipy == 1.2.1
- pyaml == 19.4.1, yaml == 0.1.7
- opencv == 4.2.0
- pillow == 6.1.0
- scikit-learn == 0.20.4
- matplotlib == 3.2.1
Training (Available soon)
- Download VDS dataset and edit 'data_dir', 'validate_dir' in default_config.yaml to the desired path
- Change the 'mode' to 'train' in default_config.yaml
- Run python main.py
Testing
- Download VDS dataset sample (test_set, test_hdr) from upper link and edit 'test_dir' in default_config.yaml to the desired path
- Download pretrained weights from upper link and place the weights in the desired path and edit 'model_dir' in default_config.yaml
- Change the 'mode' to 'test' in default_config.yaml
- Run python main.py
Acknowledgement
The code for the coBi loss is folked from the Contextual_loss_pytorch
The code for the adam cent is folked from the Gradient Centralization