Constructing a 3D object model from multiple visual features

This work aims at building 3D graphics models of objects using multiple visual cues. Different materials and shapes may yield different visual cues such as corner, pattern, highlight, and contour on objects. These cues may have similar appearance in the images, extracted as edge, peak and shading. To obtain a correct shape, different shape recovery methods must be applied to those features separately. We rotate an object and take a dense image sequence. In each epipolar plane image of rotation plane, we classify visual cues according to their image characteristics and motion behavior, and integrate shapes from different algorithms. Features are classified when they pass through a series of modules which first identify distinct features and then reasoning ambiguous features in reference to the identified ones. A carefully considered strategy is proposed and experiments on real objects.