This is my "v6" (1.6) s/w running on a random database of around 1100 images.
The image files are a mixture of gif & jpeg, and some have been deliberatly "pixelised" to make recognition a bit harder.
As of yesterday, the s/w was getting an error rate of around 4%. But when it mis-locates a face it truly is a spectacular failure. (E.g. face inside cat's ear, in the hair of the person otherwise located, etc).
The s/w is designed to be very lightweight (around 50 KB of arm code) and process at 10-20 fps on 500 MHz processors.
It will eventually have many bells and whistles, including tracking specific individuals, deciding which film star you look like, and scanning through videos looking for certain scenes (e.g. frames with a certain list of actors, etc).
lol looks like clown drawings
zerosix1786 2 years ago