Model Reconstruction for a Virtual Interactive MERLIN

A system that constructs a three dimensional model using two dimensional images taken from multiple viewpoints is presented. The images used as input were obtained by filming an object, which was being rotated on a turntable, against a dark background. The modelling process begins with the extraction of “silhouettes” from the input images. These silhouettes are used in conjunction with the a camera model to construct a volumetric representation. A sequence of “depth maps” are used to add surface concavities to the model. Following this, the surface points are found and used in the final step of the process, which entails the fitting of a triangular mesh to the model. The system is based on the approach suggested by Niem [Nie94], but includes several optimisations and extensions. One such extension is the ability to automatically calibrate the camera, so that no previous information for camera parameters is assumed. The camera parameters are calculated from the image sequence [Fit], whereas Niem’s system requires a known test pattern to precalibrate the camera. Another significant modification is the inclusion of a new silhouette extraction technique, which is able to perform accurately despite the presence of obscurring shadows in the original images. Such shadows are possible if the object is not filmed under the controlled lighting conditions that Niem’s system assumes.