I think everybody LOVES markers. I know there have been some efforts in the past but I thought why not revisit it. I came up with the following after some quick research with Gemini.
-
A marker is always associated with a scene. (a reasonable assumption)
-
Now for the source marker, we store a series of phashes (how many per second is a parameter, more frames = high accuracy but more storage and compute when we do the actual matching)
-
On the user device, all that’s left is to find where this marker slice fits into the local copy of the scene.
-
Now here’s the input I got from Gemini. There’s an algorithm called Dynamic Time Warping which precisely matches or unmatches clips in the main scene, is resistant to different frame rates and is not the bottleneck of the entire solution.
-
The bottleneck would be the generation of phashes of the local copy of the scene and we have to do that only once per scene, not once per marker.
I haven’t dug deeper but apparently a 40 minute 4K video, it would be around 20-40 seconds of cpu time.