H.Y. Ip and K.L. Chan (PRC)
Volumetric modeling, photo-consistency, specularity, voxel mask
This project aims to develop a system that can reconstruct photorealistic 3D model of object automatically from a multi-view image sequence. The whole system consists of four major steps: camera calibration, volumetric modeling, surface model formation and texture mapping. A conservative volumetric object model is first reconstructed based on the shape-from-silhouette approach. Conventional volumetric modeling algorithms for refining the model assume that the object surface is Lambertian but many real objects exhibit both diffuse and specular reflections of illuminated light. We propose a novel volumetric modeling algorithm that can handle the co-existence of both diffuse and specular surfaces. Our algorithm also differs from others that the model is reconstructed by aggregating evidence in the voxel space via the use of voxel mask. We select some real objects for modeling and compare our volumetric modeling algorithm with the well-known Space Carving Algorithm. Reconstructed object models are presented, together with the model reprojection accuracy, to demonstrate the superior performance of the new 3D model reconstruction scheme.
Important Links:
Go Back
