Facial age estimation from web photos using multiple-instance learning

One of the main bottle-necks in traditional facial age estimation is the lack of training sample problem. The rapid development of Internet provides us new chance to solve this problem. Unlimited number of facial images with their age labels can be collected through web mining technique. These images together with their surrounding text description make up the simplest cross-media data representation. In this paper, we model this problem within a Multiple Instance Learning (MIL) framework, and a novel algorithm named Witness based Multiple Instance Regression (WMIR) is proposed. The "witness" faces in the group photos are found together with their age label and confidence. A probabilistic weighted Support Vector Regression (pw-SVR) method is designed to utilize these cross-media data for learning a more robust age estimator. Experimental results upon both the synthetic data and real web data have verified the advantage of our algorithm compared with other related methods.