A representation of a three-dimensional object is autonomously learned from a sequence of the rotating object. The representation consists of single views in form of graphs and is achieved by performing a segmentation-based tracking of the object. First we apply a segmentation algorithm which is based on gray level values. This provides the location o£ the object in the images and a rough shape of it. Then we position landmarks on the object in the first frame of the sequence. These landmarks are tracked throughout the sequence on the basis of Gabor wavelet responses and guided by the segmentation result. During rotation landmarks are lost and new landmarks are added when object parts vanish or come into sight, respectively.