Generating Dense Point Correspondence Ground-Truth across Multiple Views

A ground truth dataset representing dense point correspondences across multiple views is useful in evaluating algorithms in a range of multiview geometry applications. Common datasets sparsely label point correspondences across views by either hand-marking corresponding points or by using identifiable fiducials in the scene. A few datasets feature dense correspondences but these have significant drawbacks: (i) methods using camera calibration and a laser scanner result in significant correspondence errors due to inaccurate depth estimates, (ii) methods using structured light can suffer from imaging artifacts or limitations. In addition, most of these datasets have only limited horizontal translation, not depicting wide-baseline challenges such as occlusion and intensity variations. We propose a probabilistic framework using a structured light approach where the likelihood of pixel correspondences is measured. We show that a logarithmic representation of ratios of images is the proper domain to assess the likelihood that an image pixel corresponds to a given illumination pattern. The result is a probabilistic dense correspondence map which can be used for evaluating multiview algorithms. We have created a dataset containing 13 high resolution images of a complex scene taken from distinct views which is lit using three different projectors. The resulting multi-view correspondence will be made available for public use.