aitorshuffle / ntire2018_adv_rgb2hs

Repository with the code related to the adv_rgb2hs team's submission to NTIRE2018 spectral reconstruction challenge

Requirements

• python 3.6
• Python packages: pytorch (torch, torchvision), skimage, spectral, colour, numpy, h5py, PIL, dominate, scipy, hdf5storage, tqdm, joblib

Execution instructions

• Download the code

$ git clone https://github.com/aitorshuffle/ntire2018_adv_rgb2hs.git
$ cd ntire2018_adv_rgb2hs

• Place the input RGB images to be processed in the datasets/icvl_ntire2018/NTIRE2018_Test_Clean/ and/or datasets/icvl_ntire2018/NTIRE2018_Test_RealWorld directory

• Run the rgb to hyperspectral conversion:

  • Make the execution scripts executable:

   ntire2018_adv_rgb2hs$ chmod 777 ./scripts/test_ntire2018_adv_rgb2hs_Clean.sh
   ntire2018_adv_rgb2hs$ chmod 777 ./scripts/test_ntire2018_adv_rgb2hs_RealWorld.sh
  

  • Run the execution script for each track:

    • Clean track:

    ntire2018_adv_rgb2hs$ ./scripts/test_ntire2018_adv_rgb2hs_clean.sh 
    

    • RealWorld track:

    ntire2018_adv_rgb2hs$ ./scripts/test_ntire2018_adv_rgb2hs_clean.sh
    

• Output results will be generated in:

  • Clean track: results/29
  • RealWorld track: results/34 Each of these contain an images directory, with the predicted hyperspectral mat file in the required format and one RGB image triplet per test image:
    • TEST_IMG_NAME.mat: predicted hyperspectral mat file
    • TEST_IMG_NAME_real_A.png: input RGB image
    • TEST_IMG_NAME_fake_B.png: predicted hyperspectral image rendered as sRGB
    • TEST_IMG_NAME_real_B.png: Ground truth hyperspectral image rendered as sRGB. Only makes sense for validation. At test time There will also be a index.html web page rendering all the mentioned rgb triplets.