For an in-house media tool build: pulls the good frames out of a video on your own machine, for the price of electricity. A local tool that takes a video and returns a set of distinct high-interest frames as images.
We built this one for ourselves: an in-house media tool build. It earns its place here on the same evidence as any client project.
Choosing thumbnails means scrubbing through footage by hand, which is slow and gets skipped when there is a backlog. The cloud services that do it charge per video and require uploading the footage, which is a non-starter for anyone holding client material they are not allowed to hand to a third party.
A local tool that takes a video and returns a set of distinct high-interest frames as images. It finds scene changes, looks for faces, text and people in the frame, and reads the audio for the moments where something is actually happening. It runs on an ordinary processor with no graphics card, and nothing is uploaded anywhere, so footage under a confidentiality agreement never leaves the building.
This one is built but not live yet, so there is no result to measure. The day it launches and the numbers come in, they go here, dated.
Build record: 55 tracked changes over 4 days, 21,751 lines of working code, 19 automated test files. Every number here comes straight from the project's build history.