Obtaining the Shape of a Moving Object with a Specular Surface

Recovering the 3D surface of an object from shaded images has been long investigated in computer vision. A technique of using the shading variance due to object motion has been recently studied and attracts attention [3, 4, 8]. In this paper, we also concern ourselves with this technique because it allows us to stay with the simplest possible setup. The basic scheme of the technique is to employ the framework of multi-view stereo, which first requires correspondence of several sample points to recover the extrinsic camera parameters, and then to exploit the shading variance due to object pose for dense matching. Notwithstanding the recent advances, one of the limitations of the previous techniques is concerned with the object surface, which is commonly assumed to have Lambertian reflectance. That is, if the surface takes on specularities it will be a problem for the first stage to accurately compute point correspondence. At the second stage, in which the linear subspace constraint [7] is imposed, specular reflections will be even more problematic as they produce non-linear highlights [6]. The motivation of this paper is to overcome such a limitation at each stage in the scheme. This paper is most relevant to the work of [3] and [4] in that we also investigate the linear subspace constraint. However, it differs from them in the challenge that we propose techniques to deal with specular components both in feature correspondence and dense depth search. Moreover, as a direct application we show that we can acquire a linear image basis of target object by using the consequence of dense matching [5] and removal of specular components. Our assumption is that the target object is illuminated by a distant unknown light source. The surface may or may not be textured, and is approximated by a Lambertian component plus a specular component but without shadow.