Custom Search

Monday, October 25, 2010

Jonathon Phillips' Talk

Jonathon Phillips  delivered a talk today on David Jacob's Seminar: Using Challenge Problems to Advance Biometrics

He is a researcher from NIST and from his talk, I strongly felt that their research has a kind of Physics-style. In which way, his group has made a tremendous contribution to the area of face recognition. 

His talk mainly about several face recognition challenges. He said there are four main problems in this area: Age, Expression, Illumination and ( I forgot...). I used to think face recognition is a well-solved problem, as it seems to be the most successful application of Computer Vision into business. However, from Jonathon, the face recognition problem is far away from being solved. Human can easily recognize a friend under several severely bad conditions, such as low illumination. However, the performance of face recognition program drops extraordinarily. Jonathon's team conducted an experiment of collecting good, acceptable and bad images containing static faces to evaluate several face recognition algorithms' performance. 

I think future Computer Science will divided ( or maybe has already divided) into two parts, just like the Physics now: Experimental Computer Science and Theoretical Computer Science. For example, Machine learning belongs to Theory and Computer Vision is largely an Experimental one. Do more experiment rather than just analyze data on machines will be the main trend, which is my belief too.


  1. I don't know too much about CS. What brings up to my mind is how neurons in human brain recognize faces. I know there is something called neural networks and I am not quite sure if that belongs to machine learning field. Ideally, if machine can mimic what human beings do, which, to some extent are always the optimal solution to any realistic problems. More statistics? maybe. That's just my naive thought as an outsider to CS

  2. @ xboy: Artificial Neural Network is basically regarded as a part of ML. Though it is lack of solid mathematical basis, but still it is very useful.
    However, it is widely accepted that ANN has very little connection with actual human neural networks... ANN ancestors just got inspired from human neural networks...


Custom Search