This is the implementation of CVPR 2018 Oral paper "Revisiting Deep Intrinsic Image Decompositions" by Qingnan Fan et al..
Check my homepage for the paper, supplemental material, slides and the other info.
Our codes are implemented in Torch framework. Some self-defined layers are included in the 'compile' folder. Before testing or training the models, you need to install the latest torch framework and compile the *.lua under the nn module, the *.cu and *.h under the cunn module, and adam_state.lua under optim module.
To be specific, the lua files have to be put in ./torch/extra/nn/ module directory, and editing init.lua file to include the corresponding file. Finally run the command under nn directory, luarocks make ./rocks/nn-scm-1.rockspec. Then nn module will be independently compiled.
Similarly for the cuda files, they need to be put into ./torch/extra/cunn/lib/THCUNN/ module directory, then editing the ./torch/extra/cunn/lib/THCUNN/generic/THCUNN.h to include the corresponding files, and finally run the command luarocks make ./rocks/cunn-scm-1.rockspec under ./torch/extra/cunn folder to compile them.
Accordingly, the adam lua file has to be put in ./torch/pkg/optim, edit the init.lua file and then run luarocks make optim-1.0.5-0.rockspec.
The training data for IIW can be directly obtained from their official project page. The training data for the guidance network of IIW data is L1smooth results, which is shared from this link. Note the IIW label is transfered by us into txt format for training, we include these IIW txt label files into this link. As for MIT and MPI-Sintel dataset, whose data sources may be inconvinient to obtain, we also include them in the last link.
Run gen_trainData_MPI.m to crop and flip the MPI images to generate our training data. The input images to our network have to cropped to image size of even number as we downsample the intermediate feature maps by half.
The sparse label of IIW dataset is in json file format originally, we extract the useful information and transfer them to txt file. You can either transfer the labels using transferIIW.m or directly use the transfered files in our released dataset.
The training and testing file lists are included into the .txt files in the dataset folder.
The trained models are all in the netfiles folder.
To train the models from scratch, run the files in the training folder.
Since we levearge two-fold cross validation for the MPI dataset, training on both the train and test split of MPI data is required.
To test our trained model, run test_*.lua files, while to evaluate the numerical errors, run compute_*_error files.
You can use our codes for research purpose only. And please cite our paper when you use our codes.
@article{fan2018revisiting,
title={Revisiting Deep Intrinsic Image Decompositions},
author={Fan, Qingnan and Yang, Jiaolong and Hua, Gang and Chen, Baoquan and Wipf, David},
booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition},
year={2018}
}
If you find any bugs or have any ideas of optimizing these codes, please contact me via fqnchina [at] gmail [dot] com