semillerocv / hands-on-computer-vision Goto Github PK

The problem is that a RAM overflow is generated when processing the spectral image. The easiest solution that worked for me to run the entire notebook is to use fewer channels from the time the spectral angles are calculated onwards.

For example: result_sam = spectral.algorithms.spectral_angles(cube[600:,::,:,:], materials_matrix). Or simply when loading the cube use maximum 600-700 bands.

When running the Notebook some problems are generated when executing the last cells, but they are due to some parameters received by the remap function of cv2 inside the undistortRectify function.

I have a slightly different version of the code that generates the same output as the original and it works in practice.

def undistortRectify(imageR, imageL, stereoMapL_x=stereoMapL[1], stereoMapL_y=stereoMapL[0], stereoMapR_x=stereoMapR[0], stereoMapR_y=stereoMapR[1]):

    # Undistort and rectify images
    undistortedL= cv.remap(imageL, stereoMapL_x, stereoMapL_y, cv.INTER_LANCZOS4, cv.BORDER_CONSTANT, 0)
    undistortedR= cv.remap(imageR, stereoMapR_x, stereoMapR_y, cv.INTER_LANCZOS4, cv.BORDER_CONSTANT, 0)


    return undistortedR, undistortedL

With the current documentation and hints I suffered a bit realizing the notebook, it would be nice some more help about:

• Pixel Purity Index: https://pysptools.sourceforge.io/eea.html?highlight=ppi#id6
• Fully constrained least squares (FCLS): https://pysptools.sourceforge.io/abundance_maps.html?highlight=abundance_maps#pysptools.abundance_maps.amaps.FCLS
• SPICE: https://github.com/GatorSense/SPICE
  • Params is initialized with the SPICEParameters() class.
  • inputData needs to be transposed beforehand so that it matches the input expected by SPICE().
  • The type of inputData needs to be changed to double; otherwise, the function that implements the Quadratic Programming (QP) solver using the Goldfarb/Idnani dual algorithm CRASHES.

We have to modify that equation, otherwise the function below to implement that equation does not work. Related documentation:

• https://pypi.org/project/rindcalc/1.2.2/
• https://davinci.asu.edu/index.php?title=msavi2

I noticed that the data loading, visualization, dataset and dataloader sections seem to be based on a Kaggle notebook (link). However, some important topics, such as the explanation of the UNET model and the training, evaluation and inference loops, are not detailed.

If it is OK with you, I could propose an improved version that includes additional documentation and references to tutorials that could be useful to develop the notebook more completely.

It would be nice to add %%capture in some cells to make the notebooks cleaner.

• Sesiones/Sesion2/Sesion2_1_Demosaicking.ipynb
• Sesiones/Sesion2/Sesion2_2_CameraPipeline.ipynb
• Sesiones/Sesion2/Sesion2_4_ConvolutionFiltering.ipynb
• Sesiones/Sesion4/Parte-1/Sesion4_1_HyperSpectral.ipynb
• Sesiones/Sesion4/Parte-2/Notebooks/Classification.ipynb
• Sesiones/Sesion4/Parte-2/Notebooks/Exploracion de datos.ipynb
• Sesiones/Sesion4/Parte-2/Notebooks/TargetDetection.ipynb
• Sesiones/Sesion4/Parte-2/Notebooks/Unmixing.ipynb
• Sesiones/Sesion5/Parte 1/Sesion5_1_PassiveDepth.ipynb
• Sesiones/Sesion5/Parte-2/experimental_passive_depth.ipynb
• Sesiones/Sesion6/Parte-1/structured_light.ipynb
• Sesiones/Sesion6/Parte-2/ToF.ipynb
• Sesiones/Sesion8/DeepLearning_Segmentation.ipynb