Processing - Zarand3r/deepStitch2 GitHub Wiki

Optical Flow

  • convert_using_flownet.py
  • convert_using_dali.py

Kinematics

  • preprocess.py
    • splices kinematics into demonstrations
    • creates the demonstrations labels file
      • splits samples and corresponding kinematics into demonstrations
      • timestamp resets after each demonstration
  • interpolate_kinematics.py
    • uses the demonstrations labels file to interpolate kinematics for each time segment
    • TODO: make and inclusive (store_true) option to determine value of end_index (-1)

Skills Labels

  • split.py
    • generates training data for skills assessment, writing to utils/settings.label_directory
    • 2-letter segment (e.g. CD)
      • Skills label of interest is specified as argument
      • Samples from utils/settings.data_directory with positive skills label are copied into positive_CD
      • Samples from utils/settings.data_directory with negative skills label are copied into negative_CD
    • 1-letter segment (e.g. B)
      • Skills label of interest is specified as argument
      • For each demonstration, splices a 1 second window (by default) sample around time point B
        • Samples with positive skills label are written to positive_B
        • Samples with negative skills label are written to negative_B
      • Need to then generate optical flow for this segment by supplying folder path to convert_using_flownet.py
        • This will generate the optical_flow folders inside positive_B and negative_B
        • Need to manually replace these two folders with their respective optical_flow subfolders because the training scripts assume the rgb+flow samples are in the first level of the positive_B and negative_B folders
          • TODO: make this operation automatic in the local convert_using_flownet.py script
      • Need to also generate kinematics for each sample.