We present an appearance-based technique for image characterization and retrieval. Our method is translation/rotation and scale- invariant and encodes the significant data in the image without using any segmentation. It is also very well suited to small viewpoint changes and is robust to noise and occlusi on. We present several retrieval examples in large benchmark databases, inclu ding face databases and a database of 3D objects, for which the method reaches an ideal recognition rate.