DeepIlluminance: Contextual Illuminance Estimation via Deep Neural Networks

We release the Caffe code of DeepIlluminance.

Reference

If you find our paper and repo useful, please cite our paper. Thanks!

@article{Zhang2019illuminance,
    title={DeepIlluminance: Contextual Illuminance Estimation via Deep Neural Networks},
    author={Zhang, Jun and Zheng, Tong and Zhang, Shengping and Wang, Meng},
    journal={arXiv preprint arXiv:1905.04791},
    year={2019}
}  

Prerequisites

The code is built with following libraries:

• Caffe
• Python 2.7
• Matlab 2016b

Data pre-processing

• We have trained on ColorChecker and NUS-8 datasets. Please refer to ColorChecker and NUS-8 datasets for the detailed guide of data generation. Basically, the processing of image data can be summarized into 3 steps:
• Sample image patches containing both bright and dark pixels (refer to search_patch_neighbor.py)
• View the gamma correction patches (refer to gamma.m)
• Generate LMDB files (refer to create_data_lmdb.sh and create_lmdb.py)

ColorChecker pretrained models

Here we provide the pretrained models on ColorChecker for fine-tuning at WeYun: https://share.weiyun.com/50GG5jx or GoogleDrive: https://drive.google.com/open?id=15tvz2DzlCQi3VgOpghGtkCwcf1giorWE.

Testing

For example, to test the downloaded pretrained models on ColorChecker, you can run python context_network/trained_model/test.py to get the output of the contextual network. Then, run python refinement_network/trained_model/test.py to get the final estimation result.

Training

We provide codes to train DeepIlluminance network with this repo: For the contextual network: run python ./context_network/train.py For the refinement network: run python ./refinement_network/train.py