Skip to main content

CCTV Pipeline from nbdev Docs

Summary

  1. Video List Creation
  2. Frame Extraction
  3. Azure OCR
  4. Blurring Frames
  5. Extracting Distance Values from Frames
  6. Extracting Condition Codes from Frames
  7. Split Data into Training and Testing
  8. Run Multiple CCTV Pipeline Modules in Sequence

Notes on Summary

It looks like this summary is missing some steps that exist in CCTV Workflow, namely steps 2 and 3 in that other workflow are not present in nbdev docs: 2) "Extracting metadata from metadata DBs" and 3) "Import supplementary data"

My best guess at this point is that I need to follow a flow that is a combination of the two. After lunch, I will try to follow the two flows combined (I already have the meta extracted). I will talk this over with Sudhir.

Sudhir notes on 2 processes

  1. Paper 1 - Analyzes videos (documented in docusaurus and nbdev)
  2. Paper 2 - AKA Light GBM - Predicting defect locations in GIS (documented in docusaurus ONLY)

DNV 2024 deliverable is just Paper 2 flow for now.

This means I don't need to deal with frame extraction and all that for the current deliverable.

Paper 2 is a 1-to-1 with our docs folder "lightGBM" (technically M-to-1, but we are currently only interested in LGBM as the "best" model).