hermes-config/skills/mlops/computer-vision-inference/references/video-content-renaming.md

Video Content Detection + Renaming — Production Script

A working YOLO-based pipeline for detecting video content from frames and renaming generically-named recovered video files (FILE000.MOV pattern).

Script

Saved at /tmp/rename_videos.py and /tmp/rename_videos2.py on rayserver.

Workflow

1. List all files in the target directory starting with "FILE" (generically-named recovered files)
2. Extract a video frame at 3 seconds using ffmpeg; fall back to 1 second for short clips
3. Run YOLOv8n on CPU for content detection
4. Build descriptive filename: {people_count}_people_{object1}_{object2}.mov
5. Handle filename collisions with _1, _2 suffixes
6. Files with no detections → scene.mov

Key design decisions

• CPU inference — Viable for 80-100 files at ~3-5s/file. GPU not needed for this scale
• ffmpeg single-frame extraction — Much faster than processing the entire video
• 3-second offset — Early enough to avoid blank intros, late enough to have content
• startswith("FILE") guard — Makes the script idempotent (won't rename already-named files)

Performance reference

• 80 files: ~4-6 minutes total (i7-10700K, CPU)
• ~60 files get descriptive names, ~20 remain as scene (blank/dark)

Full script (80-file version)

See /tmp/rename_videos.py and /tmp/rename_videos2.py for the two-phase version that handled 80 files.

Edge cases

• .3GP files are valid video containers — include them
• Short clips (<3s): ffmpeg clips to available duration, produces a valid frame
• Broken headers: ffmpeg hangs if file is truncated — always set timeout=20
• Collisions: use while os.path.exists(dst) loop with incrementing counter