mmiskatul/pothole-depth-scan ? reverse-engineered prompt

Build me a Python tool that takes a road video and finds potholes, then colors each one by how deep it seems compared with the road around it, from green for shallow to red for severe.

I want it to save an annotated MP4 with clear overlays, small label chips showing a stable ID, severity, depth score, and confidence, plus a simple on screen HUD with frame count, FPS, active depth method, and counts for each severity level. It should also be able to export a CSV with per frame measurements.

Please make it work with a trained YOLO pothole model file called best.pt, and support both a neural depth option and a simple shadow based option for flat or aerial footage. Add command line options for source video, confidence, image size, ROI masking, start frame, display mode, depth preview, output CSV, and GPU if available. If you need to check current docs for any libraries, go ahead and look them up online.