I think it’s high time we admit it; despite all the work that has been done (and all the complex maths used) on face recognition systems, the performance of these algorithms and systems are abysmal. They certainly do well in lab conditions (error rates of maybe 1 in 50) but in real life conditions, they suck!!!
From the approximate figure mentionned above (1 in 50), that would seem good but it means that if 1000 people were to pass by, we would have 20 errors and in 10000, 200 errors and this is very roughly what we will have to face in real life conditions. Yet, this is not to mention variation in lighting, angle of faces,…
So the question that is to be asked is: are we proceeding the right way? Are our algorithms basically flawed? Are our basic assumptions about recognition wrong?
I think we ought to do a total rethinking of way in which we have been proceeding and try to see what is it that humans use to recognise faces. Human seem to rely on many things like colour of the skin, some oddities of a specific face,… to perform instantaneous recognition.
The other side to it is are we being fair in requesting a 100% recognition rate from computers based only on faces. Is that feasible? Do humans rely only on faces to recognise a person or is there an array of things – “I thought it was Jack but he’s way too tall to be Jack.” Isn’t that something we say a lot?
I’ve put up a lot of questions here coz I sincerely think that we need to rethink our algorithms. It simply can’t be that we’ve been doing research for about 20 years now and our results are sooooo bad!!!