Presented here is a new algorithm for finding the positions of features in images of samples from particular object classes, such as human faces. Existing algorithms that address this problem mostly deal only with image variations resulting from simple translation in the image plane, as well as differences of objects in the classes, by searching for features across the image plane. In our new algorithm, larger classes of image variations, including those resulting from object rotation in 3D space and scaling (i.e. translation in depth) are handled, in addition to image plane translation. In order to do this, we develop a new Kernel-based MAP (maximum a posteriori) estimation technique using Gaussian distribution in a potentially higher dimensional space to model the relationship between images and feature positions. Experimental results of facial feature extraction in images of human faces taken from varying viewing directions and from varying distances demonstrate the superior performance of the new method relative to that of existing algorithms.