Week 03

Nov 14

• Get labels for 39 videos
• Submit job train on 39 videos to compute canada (expected runtime = 2 hour) (get the histogram of defect classes with and without ND)
• Submit job train on 33 + 39 videos to compute canada (expected runtime = 4.5 hours)
• Submit prediction run on images from 39 videos using model trained with 33 videos
• Submit prediction run on images from 33 videos using model trained with 39 videos
• how to sample low count defect (B)

Nov 18

• I was able to fix the algebraic loop when I switched the sh2 solver and ph solver from the newton-raphson algorithm to the collimator discrete integrator. The BSM2 people decided to use Newton-raphson instead of doing it by differential equation because they wanted to reduce the stiffness of the ADM1. I will switch to continuous integrator whether the outputs are better than using the discrete integrator.

• Outputs from discrete and continuous integrator approach are similar to each other. However, both outputs are different from the Matlab outputs.

• I believe there are errors in my pH solver block. I will need to fix the ODE in the pH solver block

• CCTV dataset:

  • Please make sure that the results from the runs so far are documented (people.gqc.com?)
  • Please write up the algorithm for the labeling of frames using the continuous defect information in the database.
  • Put that in the people.gqc.com
    • Link to Current Image Labels processing algorithm
    • Link to additional algorithm to deal with access database and continuous defect
  • modify the existing code to implement that.
  • Use one video for testing that is randomly or manually selected. Use the rest for training and validation (80 20-first straight and then add cross-validation). Then calculate the metrics for that video.
  • Repeat preceding step for several test videos. Pick the run with the worst results and then analyze the results