We present a novel multi-view stereo method designed for image-based rendering that generates piecewise planar depth maps from an unordered collection of photographs.
First a discrete set of 3D plane candidates are computed based on a sparse point cloud of the scene (recovered by structure from motion) and sparse 3D line segments reconstructed from multiple views. Next, evidence is accumulated for each plane using 3D point and line incidence and photo-consistency cues. Finally, a piecewise planar depth map is recovered for each image by solving a multi-label Markov Random Field (MRF) optimization problem using graph-cuts. Our novel energy minimization formulation exploits high-level scene information. It incorporates geometric constraints derived from vanishing directions, enforces free space violation constraints based on ray visibility of 3D points and 3D lines and imposes smoothness priors specific to planes that intersect.
We demonstrate the effectiveness of our approach on a wide variety of outdoor and indoor datasets. The view interpolation results are perceptually pleasing, as straight lines are preserved and holes are minimized even for challenging scenes with non-Lambertian and textureless surfaces.