video-inventorylisted

# Video Inventory Turn a walkthrough video into a priced inventory: extract frames, read what is actually written on the objects, identify each one, look up the resale value, and hand back a report. The hard part is not the extraction. It is reading a book spine or a serial number off a handheld 832×464 phone video, and then being honest about what you could not read. ## Pipeline ### 1. Probe the source ```bash ffprobe -v error -show_entries format=duration,size \ -show_entries stream=width,height,codec_name \ -of default=noprint_wrappers=1 "VIDEO" ``` A WhatsApp video is typically 832×464. Resolution is the binding constraint, not the number of frames. If spines or serial numbers are in scope, say so up front, and say which close-up photos would settle the doubt. ### 2. Extract and rank by sharpness ```bash S=<skill_dir>/scripts/frames.py python3 $S extract "VIDEO" work/all --fps 6 python3 $S sharp work/all --per 1.0 --fps 6 ``` `sharp` prints, for each second, the sharpest frame (Laplacian variance). **Always work from those.** In a handheld pan, sharpness between neighbouring frames varies by a factor of 3 to 5, and that factor alone decides what is legible. ### 3. Map the scene Read 6 to 8 frames spread across the duration, roughly one every 3 seconds, to work out how many shelves, zones or distinct shots there are. Do not read all 150 frames. It adds nothing and burns the context you need for the details. ### 4. Zoom to read This is the step that decides