Fitting a Morphable Model to Pose and Shape of a Point Cloud书.pdfVIP

Fitting a Morphable Model to Pose and Shape of a Point Cloud David C. Schneider, Peter Eisert Fraunhofer Heinrich Hertz Institute, Einsteinufer 37, 10587 Berlin, Germany {david.schneider,eisert}@hhi.fraunhofer.de Abstract models have been used most successfully in the do- main of the human head and face, where they are The paper addresses the task of ?tting a morphable called morphable head models [9]. A wide number head model to a dense, unstructured and untextured of applications in computer vision and graphics can point cloud. The problem is typically approached be driven or supported by morphable models, in- in a multi-step process, comprising of generic non- cluding, for example, face recognition [3, 10], face rigid registration, conversion to the model’s topol- tracking [17, 18] and several varieties of 3D recon- ogy and ?tting of the model. Here, a direct approach struction, e.g. from uncalibrated video [12] or from is proposed where the morphable model is ?tted stereo images [2]. to the point cloud itself in an optimization process In order to ?t the morphable model to the shape following the Iterative Closest Points framework. of a human head—i.e. to compute the model param- Along with shape parameters, rigid pose and scale eters that describe the head—the target head must are estimated explicitly which leads to better ?tting be represented in a way that is semantically consis- results than shape estimation alone. A compact for- tent with the model: The target must have the same mulation of a cost function is proposed, applicable number of vertices as the model and these vertices also in the case of a model comprising of multiple, must be placed at semantically equivalent locations. independent sub-models each of which describes a If, for example, the vertex with index K describes region of the face.