# python-opencv tutorial
https://opencv-python-tutroals.readthedocs.io/en/latest/py_tutorials/py_tutorials.html

# conda env support in jupyter notebook
# conda install nb_conda // already in yml

# select conda env for jupyter notebook
conda env create -f cv_project.yml
conda activate cv_project

# download the datasets 
python3 demosaicing/download_kodak.py --output-dir=data/kodak

# matlab bug: duplicate libomp loaded
export KMP_DUPLICATE_LIB_OK=True

Todos

• read more on how ratio images are important
  • read proofs in c2b paper
• how to choose the subsampling mask
• what is trade-off between spatial resolution / #patterns
  • challenging casses
  • depth discontinuity
  • texture discontinuity
  • resolution
• ratio image vs. intensity images
  • motivation: ratio images do not have texture
  • RED does a lot better in ratio space ~dB increase in perf
  • yet to test: performance carry over to disparity reconstruction
• denoiser
  • re-train on noise characteristic of c2b camera
• do denoising in another domain!
  • optimized variable should be albedo, disparity, denoised image etc.
• end-to-end optimization
  • think about ways to regularize disparity etc.
  • relationship between ratio images and disparity/phase
  • probabilistic formulation or alternating optimization
• do zncc on optimized code
• think about fast algorithm for video decoding
• matrix inversion lemma on quadratic update! to simply
  • See if can use simplification in DeSCI paper here
• optimization
  • decreasing noise level
  • adaptive rate, lr, gamma, etc.
  • termination condition (insufficient update terminates the optimization)
• do hdr, hyperspectral imaging with c2b as well, joint optimization etc.
• denoiser
  • trained on c2b images, might not be that important as Gaussian seems to be an OK noise model.
• consider geometric perspective paper and use it to design better coding schemes catered to the c2b camera
  • noise no longer gaussian, noisy measurement not in an ellipsoid anymore, since noise no longer gaussian.
  • can we optimize for geometry in measurement space to design optimal code, i.e. coding curve that satisfies the 3 desideratas
• note noise characteristic of S images is dependent on the reconstruction used to create them