Probabilistic Modeling for Detection and Gender Classification

2. Related Works

In face image analysis, determining gender of face from facial trait classification is still a challenge task. The greater part of publications of gender classification highlights the state-of-the-art in general trait classification. Trait learning is done by using spatially global feature representations such as templates (Kim, 2006; Gutta, 1998), principal components (Moghaddam, 2002), independent components (Jain, 2005), or image intensities directly (Baluja, 2007). While most approaches utilize intensity data, 3D information may also improve sex classification (O'Toole, 1997). In this context, much works have explored the capacity of different machine learning such as neural networks (Gutta, 1998), SVMs, and boosted classifiers (Baluja, 2007). Recently, trait classification based on local features (Kadir, 2001; Lowe, 2004; Mikolajczyk, 2002; Mikolajczyk, 2004; Yu, 2009) has emerged, using local regions (BenAbdelkader, 2005) or Haar wavelets (Shakhnarovich, 2002; Yang, 2006). Local features are principally computed from the regions of image as the values of Histograms of Oriented Gradients (HOG) (Lowe, 2004). They are image descriptors in variant to 2D rotation which have been used in many different problems in computer vision.