[PD] Skin Detection

[Mathieu Bouchard]

On Thu, 17 Jun 2010, João Pais wrote:

>> I would be very careful with your conception of skin if your going
>> visual sensing, as it would be very difficult to make a skin detector
>> that does not privilege certain skin tones.
> like in http://www.youtube.com/watch?v=t4DT3tQqgRM ?

Typically, face recognition involves using statistics about what a face 
looks like, so if the programme ships with a precomputed neural network 
and no tool to help add custom faces, chances are that it will miss a lot 
of people. I haven't tried such tools but I could guess that it would also 
miss people with unusually large beards and large mustaches,

   http://esoterisme-exp.com/Section_main/Collection4/612-esoexp30.jpg
   http://blogue.branchez-vous.com/archives/doc-mailloux.jpg

A large mustache can cut a face in two, especially with certain angles and 
lightings. A large beard changes a face so much, that while googling for 
the latter guy, the first text I saw was someone wondering "what his face 
looks like" (without beard). A photoshop speculation :

   http://blogue.branchez-vous.com/upload/2007/04/doc_mailloux.jpg

But algorithms can't come up with that kind of picture, and even from a 
human, there's always a possibility that the chin is wrong, if there's not 
enough of a hint of it on the picture.

Finally, black faces have much lower contrast, especially if the camera 
has a narrow range of intensity and is auto-adjusting to the average 
brightness of the scene. On top of that, any shininess of the skin can 
make reflections that can confuse the algorithm a lot.

  _ _ __ ___ _____ ________ _____________ _____________________ ...
| Mathieu Bouchard, Montréal, Québec. téléphone: +1.514.383.3801