I spent the weekend working on my vision server again. The face detection part is down pat so this time I had a go at the facial recognition part.

So far I have the SOM (self organising map) is successfully categorising most of the images I am feeding into it (just using test data for now) but there still seems to be quite a few of the images that I would subjectively have though were very similar even though they are being mapped to different nodes in the SOM. It’s almost working, I think I might need to visualise the output in a different way though in order to really get it.

The main issue is that there are quite a few params to tweak – map node count, image normalisation size, histogram size and average colour grid size – not to mention the training parameters for the SOM itself so I may just have some of the numbers wrong.

On a mildly related note, the Oxford mobile robotics group has released an open source version of their Fast Appearance Based Mapping (FABMAP) toolset. It’s C++ based and is apparently easy to use. If you are interested in adding mapping and navigation capability to your bot it might be worth a peek.

I am going to check it out now and I’ll write something up if it is useful.