Head Pose Classification from Low Resolution Images Using Pairwise Non-Local Intensity and Color Differences
Teera Siriteerakul, Daisuke Sugimura, Yoichi Sato · 2010
In this work, we propose novel image descriptors for identifying head poses in low resolution images. The key novelty of our method is to exploit two types of non-local metric for estimating head poses: non-local intensity difference feature (iDF) and non-local color difference feature (cDF). Unlike the existing methods that one pixel can only represent one head pose information, our proposed features are designed to capture geometry of head pose image via relative information of two-randomly picked pixels. The iDF is designed to capture relative head image regions represented by the two pixels ( e.g. one pixel represent hair while the other represent skin ) without explicitly labeling any of the pixels. On the other hand, the cDF is designed to capture information about whether or not the two randomly-selected pixels belong to the same head image regions, again, without explicitly labeling any of the regions. Our experimental results demonstrate that our descriptors using pair wise differences in intensity and color outperform current state-of-the-art for head pose estimation from extremely low-resolution images.