exadel-inc
Track faces to improve recognition
- Describe what you are trying to accomplish and why in non technical terms Track faces it received from double take (might be better to implement this on double take side) so it does not need to keep matching faces and potentially choose a lower match at some point. Once a face with good match and percentage is received, track it across a camera (or cameras e.g. a recognised face can't exist on more then one camera, and if it does, flag it for potential spoofing)
This will: -Improve recognition -Lower cpu and overhead -Implement spoofing -If the tracked face goes to a place with a small recognition box, it won't have to rely on small images but can use the previously recognised face instead.
- Describe the solution you'd like Track faces or the object attached to it to lower cpu and improve recognition. Describe alternatives you've considered Doubletake, to which i will also suggest this.
- Additional context Add any other context or screenshots about the feature request here.
This is a popular task, and many people ask why we don't have it in CompreFace. I believe CompreFace is not the right place to implement it. CompreFace is a generic solution for lots of facial recognition cases. Many of them don't need recognition from the video at all. But even if the cases need recognition from video, the requirements are always different. This is why I believe instead of extending CompreFace, different solutions should appear based on CompreFace for each narrow case. DoubleTake is one the good examples.
Hmm fair enough. But do you not agree it would improve face recognition ? If so, would it not fit within Compreface as that is it's sole task. Not criticising by the way, just thinking the above would improve recognition by quite a margin (but i might super wrong).
And you might be right, maybe a better option for 3th party addons.
It will improve face recognition accuracy. But CompreFace is not the right place for it :)
@atv2016 Hey, did you end up having any luck with tracking for faces? Any clue on how to start on it as I don't think there's a way to know if the person from the previous detection is the same person as now (especially if there are multiple people). Maybe something like tracking the xy axis of the bounding box and calculating if the movement was incremental and assuming it's the same person?
