Week 2

June 06/09/23

Experiment to find partially detected frames

1. Running the distance + defect extraction nb will result in the following cases.

note

The distance + defect extraction nb does not break if the distance text is not extracted using regex from OCR output. All it does is return null.

Nuances

1. frames with distance extracted from OCR output
   1. Other fields can have partially identified text which won't be recognized by our experiment. For example, defect annotation can be missing.
2. frames without distance extracted
   1. original frame did not have distance (this is not a problem)
   2. original frame had distance, OCR failed to recognize

danger

This experiment will not be able to differentiate between 2i and 2ii without manually inspecting the frames.

June 06/08/23

1. Running the blurring notebook
   1. Convert the nb as a py script. (Deven used nbdev_export) to create the py file
   2. Use the same groups as OCR for processing
   3. Make sure the notebook is updated to save jpgs in full resolution.
   4. Make parallel runs on MSI Server
2. Steps on the AI3 front
   1. The lookup function is being revised.
      1. When found an exact match it, it returns (word, True, jaccard_distance) as expected
      2. It is slightly tricky, when the key is not found.
      3. The reason behind this struggle is that, this was a recursive function written by ChatGPT and adding new functionality in the recursive function is more tricky than expected.
   2. Regarding generating more labelled,
      1. I haven't had the chance to take the laptop home and run the script which will label JSON sentences using Bard in peace.
      2. I am looking at alternatives which can be used directly in the office.

June 06/07/23

1. Frame extraction from all the videos with revised ffmpeg arguments
2. Move compatible-mp4 and uncompatible-mp4 videos to Received Data folder so that the video_list csv can be created.
   1. This is no longer required.
   2. The original zips were extracted again and frame extraction was run.

Frame extraction

1. Extract all frames from all the videos
2. Zip the jpgs
3. Save them to Extracted_Frames folder

Azure OCR

1. Modify the stitching command to write the jpg in full-resolution
2. Try to write to ramdisk to improve speed-up
3. Update the logic to create either a new-folder or save to a new temp file

To run OCR in parallel

1. Deven's thought
   1. Convert the nb to py file and create a CLI out of it
   2. Open n terminal windows using tmux and make separate calls manually by changing the arguments.
2. Pavan's thought
   1. Modify the shell script that Vannary used to submit n jobs in parallel on SLURM.
   2. Use the & sign to push the process to background and continue.

June 06/05/23

1. Run Azure OCR on Group 1. We need to create a new Azure Vision instance in the Azure Portal.
2. Run the blurring notebook.