Accurate depth-map refinement by per-pixel plane fitting for stereo vision

This paper discusses the refinement of sparse and noisy depth-maps to improve stereo measurements. Our method functions as a post-filter for stereo measurements, to remove outliers and interpolate the depths of invalid pixels. Per-pixel plane fitting is employed to estimate the normals of an object's surface in a depth-map. These normals provide information regarding the interpolation of depth and the removal of outliers by evaluating the directions of surfaces. In our experiments, our method successfully reconstructed a dense and accurate geometry from a sparse and noisy depth-map, even where several dozen percent of pixels were outliers and only a few percent were from the original correct geometry. This result indicates a novel method of fast stereo measurement, because dense reconstruction can be performed without stereo matching for all pixels.