Photorealistic 3D Model Reconstruction based on the Consistency of Object Surface Reflectance Measured by the Voxel Mask

Abstract

To create three-dimensional (3D) models of real scenes and objects is an old and challenging computer vision problem. Systems that can reconstruct the 3D model of object, for instances the human head or cultural artefacts, have found many applications such as virtual character animation and interactive museum exhibition. Real object models can be reconstructed automatically using active and passive methods. Object range scanning by laser or structured light are typical examples of the active methods. They often demand expensive equipment and special skill to operate. Moreover, they are not very good in modeling very glossy objects. The passive methods can acquire images of the object at different viewpoints using off-the-shelf CCD cameras (Chang & Chen, 2002). The camera is usually calibrated by taking pictures of a specially designed calibration pattern or object. The camera viewpoints can be arbitrarily selected and the camera model is adjustable. For instance, Niem (Niem, 1999) proposes a 3D object reconstruction method using a mobile camera to capture image of the object and calibration pattern simultaneously. Our system of 3D object model reconstruction consists of four major steps: camera calibration, volumetric model reconstruction, polygonal model formation and texture mapping. An overview of our system is shown in Figure 1. The ca mera calibration is to obtain the intrinsic and extrinsic parameters defining the internal camera properties and the viewpoint orientation with respect to the object. The object and the calibration patterns can be captured simultaneously. Therefore, the camera can be placed anywhere and each view can be calibrated independently. One of the popular approaches for volumetric modeling is shape from silhouette (SFS), which is to recover the shape of object from its contours. However, reconstruction of a complex rigid object from its images is a challenging computer vision problem, especially when the object exhibits large textureless surface or concave surface. Previously, we enhance the SFS-based volumetric modeling algorithm by imposing the photo-consistency in neighboring views and the aggregation of evidence in volume space via the use of voxel mask (Wong & Chan, 2004; Chiang & Chan, 2006). Although the algorithm is very good in tackling textureless as well as concave surface, it is still unable to model non-Lambertian object surface accurately. In the present investigation, we propose a novel volumetric modeling algorithm that further improves the shape reconstruction by explicitly taking into account the object surface specularity. Then the polygonal model is