Comment 62 for bug 507511

Are there any good classifiers out there that can be trained on a generic set, and then have good enough accuracy to perform well without any additional training and improvement? Surely the classifier has to be trained on the specific dataset of images first (users database, maybe not the whole, but a selected few images), and then continuasly improved by the user feedback as it is used. Mac's iPhoto has this feature and even with ongoing improvement it still gets the a lot of false positives, that are then weeded out by the user. This way it is not limited by the accuracy of any model that is trained elsewhere.

An obvious case where classifier might perform poorly, is when faces are partialy ocluded by some object. It might recognise it as another person or not recognise it at all. In which case it has to be told "No this is not object A, it is object B". I have also seen that using a number of Haar features can still yield a non face recognised as a face by similarity. So the classifier needs to be told to ignore that somehow.

An additional comment is that, as I understood from description above, the database field caters for a position of a face, not multiple faces. May I suggest to use a list of positions where faces are and what they correspond to. So have a separate classifier table, that has internal representation of people feature set and internal id of a person, which can be then tagged to the face position and id of the person.

E.g. <coordinates of the bounding box (top right and bottom left)>#id of the person: <125,125>,<175,175>#3

Admin question: has this been submited to the GSoC 2009, because I couldn't see it in the project list? Is anyone working on this? Otherwise I will throw my name in the hat to help out with this :)