Author: ZhuoYi Cai
Email: zycai@stanford.edu
Course: EE367
Project Title: HDR image reconstruction from a single short- exposure image

This code package can be used to reproduce the result for this final project. Since this is built from scratch within one and a half week time frame, we only have limited comment. Please see actual code for the details.

There are several notebooks to record the result for different experiments done during this project.

ee367FinalProject.ipynb - overfitting experiment
ee367FinalProject4.ipynb - 128x128 patches
ee367FinalProject6.ipynb - 64x64 patches

Other notebooks are the result of experiments such as using 2D convolutino vs Max pooling as downsampling or adding Max pooling on the skip connection etc. See code for details if interested.

model*.png are all model plot for different experiments.

code
├── data (We don't include data here. It is 4GB only for the subset. See Dataset below for more details)
│   ├── demo
│   ├── test
│   │   ├── gt
│   │   └── raw
│   ├── train
│   │   ├── gt
│   │   └── raw
│   └── val
│       ├── gt
│       └── raw
├── ee367FinalProject.ipynb
├── ee367FinalProject1.ipynb
├── ee367FinalProject2.ipynb
├── ee367FinalProject3.ipynb
├── ee367FinalProject4.ipynb
├── ee367FinalProject6.ipynb
├── model.png
├── model1.png
├── model2.png
├── model5.png
├── model6.png
├── model8.png
└── srccode
    ├── __init__.py
    ├── data.py
    ├── models.py
    ├── models0.py
    ├── models1.py
    ├── models4.py
    ├── nn.py
    └── util.py


Dataset:

This code package doesn't include any data for the training and validation. Please run the code in the notebook to download the data from google HDR+ dataset. Also, to reproduce the result, the data needs to be pre-processed manually to make sure the raw image and ground truth image have the same orientation.

history directory:

This directory contains several training plots as the result of various experiments.
