We noticed that our camera was recognising pictures of faces in the newspaper to focus on, so I explained how Haar cascades worked - this is the algorithm that cameras use, I understand.


Here's a diagram of some of the Haar features, superimposed on their face locations:

Here's the image from the paper, showing the two top Haar features for a face:



Turtle started copying my diagrams of the Haar features.


We decided to see whether we could sketch a face sufficiently well that the camera would recognise it and focus on it. This face looks great but didn't work.


Adding a mass of hair around this face made it work - the camera identified it as a face. But it hasn't worked in Picasa, they must have different parameters for their face detection.

I was at CVPR in 2001 and will never forget the moment when the speakers turned the video camera on the audience and we saw real-time face recognition for the first time, with a cloud of boxes appearing over our faces. The reaction of the audience (serious academics) was amazing, like a bunch of delighted children, all standing up and waving.

This post was originally on LiveJournal.