CRF-Based Reconstruction from Narrow-Baseline Image Sequences.

Given an image sequence of a scene it is possible to recover a depth map. Though multiview stereo algorithms are well-studied, rarely are those algorithms considered in the context of narrow baseline. In this paper, a practical method is proposed to generate dense depth map using a narrow-baseline image sequence. We introduce a new structure from small motion method tailored for narrow baseline which allows us to recover sparse scene structure and camera poses. In the dense reconstruction, we adopt a space-sweeping method for dense matching and a fully connected conditional random field model for depth refinement. As opposed to prior methods that guide CRF with color information alone, we creatively add depth information guidance which effectively avoids over-smoothing and bad impact from error color information. Our approach produces higher-quality dense depth results than state-of-the-art algorithms under the same baseline configuration.