Superpixel-based depth estimation for multiple view stereo

Superpixels are the small regions in which the pixels are consistent in color, intensity and texture. The superpixels contain the principal features of the particular regions, and can remain the important boundary information at the same time. It is the natural unit to describe an image, compared to the pixel which is considered as the physical unit. In this paper, we propose a novel method on superpixel-based depth map merging MVS algorithm. A Markov Random Fields model with more possible depth candidates from neighboring segments is built to acquire the depth assignment for each pixel. Moreover, in order to obtain more accurate depth information, a novel depth map generation framework with respect to multi-scaled superpixel segmentation is employed. The experimental results prove the effectiveness of our proposed method.