Reconstruction in the round using photometric normals and silhouettes. George Vogiatzis 1 Carlos Hern´ andez 2 Roberto Cipolla 3 Dept. of Engineering, University of Cambridge, Cambridge, CB2 1PZ, UK {gv215 1 , ch394 2 , cipolla 3 }@eng.cam.ac.uk Abstract This paper addresses the problem of obtaining complete, detailed reconstructions of shiny textureless objects. We present an algorithm which uses silhouettes of the object, as well as images obtained under varying illumination con- ditions. In contrast with previous photometric stereo tech- niques, ours is not limited to a single viewpoint and pro- duces accurate reconstructions in full 3D. A number of im- ages of the object are obtained from multiple viewpoints, under varying lighting conditions. Starting from the silhou- ettes, the algorithm recovers camera motion and constructs the object’s visual hull. This is then used to recover the illumination and initialise a multi-view photometric stereo scheme to obtain a closed surface reconstruction. The con- tributions of the paper are twofold: Firstly we describe a robust technique to estimate light directions and inten- sities and secondly, we introduce a novel formulation of photometric stereo which combines multiple viewpoints and hence allows closed surface reconstructions. The algorithm has been implemented as a practical model acquisition sys- tem. Here, a quantitative evaluation of the algorithm on synthetic data is presented together with a complete recon- struction of a challenging real object. 1. Introduction We propose a method for acquiring a complete 3D model of a uniform untextured object from a number of images captured under varying light conditions. The object’s re- flectance is assumed to follow Lambert’s law but a signifi- cant number of highlights are present. A sequence of im- ages of such an object is given, where the object moves in front of a fixed camera and a single distant light-source moves arbitrarily between each image capture. It is also assumed that the object can be segmented from the back- ground and silhouettes extracted automatically. Shape recovery from images is a well established com- puter vision task with two families of techniques offering the most accurate results, multi-view stereo and photomet- ric stereo (see [6] and [5] for some of the best quality results Figure 1. Reconstructing textureless shiny objects. Objects from textureless shiny materials such as the porcelain figurine shown, present a challenge for shape reconstruction algorithms. The lack of surface features makes traditional multi-view stereo very difficult to apply while photometric stereo has so far only been able to produce 2.5D reconstructions. Our algorithm is able to produce closed-surface, full 3D reconstructions of many-sided objects, from a sequence of uncalibrated images and an arbitrarily moving light-source. Here, two views of the reconstructed model are shown next to views of the porcelain object. from each method). While correspondence based multi- view stereo techniques offer detailed full 3D reconstruc- tions, they rely on richly textured objects to obtain corre- spondences between locations in multiple images which are triangulated to obtain shape. As a result these methods are not directly applicable to the class of objects we are con- sidering due to the lack of detectable surface features. An attempt was made on reconstructing such objects in [7] but the reconstructed models shown lack surface detail which is due to the regularisation enforced on the reconstructed surface. On the other hand, photometric stereo works by observing the changes in image intensity of points on the object surface as illumination varies. These changes reveal the local surface orientations at those points that, when in- tegrated, provide the 3D shape. Because photometric stereo performs integration to recover depth, much less regular- isation is needed and results are generally more detailed. Furthermore, photometric stereo makes fewer assumptions about surface texture and reflectance, which can be almost completely arbitrary as demonstrated in [5]. However, the simplest way to collect intensities of the same point of the surface in multiple images is if the camera viewpoint is held constant, in which case every pixel always corresponds to the same point of the surface. This is a major limiting factor