Combined model for detecting, localizing, interpreting and recognizing faces
Abstract
This work describes a method that combines face detection, localization, part interpretation and recognition, and which is capable of learning from very limited data, in a semi-supervised or even fully unsupervised manner. Current state-of-the-art techniques for face detection and recognition are subject to two major limitations: extensive training requirement, often demanding tens of thousands of images, and detecting faces without explicitly detecting relevant facial parts. Both of these limitations hinder the recognition task, since for a specific face the number of available examples is usually small, and because the strongest cues for identity lie in the specific appearance of facial parts. The proposed method alleviates both these limitations by effective learning from a small training set and by detecting the face through, and together with, its main parts. This is obtained by a novel unsupervised training method, which iterates phases of part geometry and part detector learning, to incrementally learn an object category from a set of unlabeled images, containing both class and non-class examples given in unknown order. We tested our method on face detection and localization tasks both in a set of 'real life' images collected from the web as well as in LFW and MIT-CMU databases. We also show promising results of our method when applied to a face recognition task.
Origin : Files produced by the author(s)
Loading...