Fast level set multi-view stereo on graphics hardware

In this paper we show the importance and feasibility of much faster multi-view stereo reconstruction algorithms relying almost exclusively on graphics hardware. Reconstruction algorithms have been steadily improving in the last few years and several state-of-the-art methods are nowadays reaching a very impressive level of quality. However all these modern techniques share a very lengthy computational time that completely forbids their more widespread use in practical setups: the typical running time of such algorithms range from one to several hours. One possible solution to this problem seems to lie in the use of graphics hardware: more and more computer vision techniques are taking advantage of the availability of cheap computational horsepower and divert graphics hardware from its original purpose to accelerate the early stages of some algorithms. We present here an almost full implementation on graphics hardware of a multi-view stereo algorithm based on surface deformation by a PDE: this algorithm tries to minimize the error between input images and predicted ones by reprojection via the surface. As it mainly works on whole images, it is well suited for graphics hardware. We show how we succeeded to bring the whole reconstruction time within minutes. Results for synthetic and real data sets are presented with computational times and compared with those of other state-of-the-art algorithms.