Obligatory XKCD [1]<p>With quite a bit of work, a research team, a development team, and quite a few servers I was able to finish a project which did face recognition at just about the accuracy you are talking about [2]. With that we increased the accuracy of the hair and scalp recognition within the OpenCV project. However, this required quite a bit of algorithmic training [3]. Also, this did not include a set of vector points drawn around the subject, just a bounding box. This is a rather difficult problem set.<p>Now, if you have seen these fun little apps which show you what your hair would look like with a new style (or what you would look like with a set of glasses, or a new dress), this uses very little actual recognition. Basically they take a picture of you within a general "person shaped bounding box" or for hair a "head shaped bounding box" and then overlay the new style on top of the area where your hair, glasses, or clothes should be. This is relatively simple (especially in comparison to actual recognition). Might it solve some of your problem?<p>[1] <a href="http://xkcd.com/1425/" rel="nofollow">http://xkcd.com/1425/</a><p>[2] This accuracy was for the entire algorithmic set. It did not include a single datapoint set like "hair". It was also highly dependent on the image quality and how the subject was positioned within the image.<p>[3] Our test image set was 50,000. Our production training image set was 1,500,000 images.