基于匹配扩散的多视稠密深度图估计

提出一种高精度的基于匹配扩散的稠密深度图估计算法. 算法分为像素级与区域级两阶段的匹配扩散过程.前者主要对视图间的稀疏特征点匹配进行扩散以获取相对稠密的初始深度图; 而后者则在多幅初始深度图的基础上, 根据场景分段平滑的假设, 在能量函数最小化框架下利用平面拟合及多方向平面扫描等方法解决存在匹配多义性问题区域(如弱纹理区域)的深度推断问题. 在标准数据集及真实数据集上的实验表明, 本文算法对视图中的光照变化、透视畸变等因素具有较强的适应性, 并能有效地对弱纹理区域的深度信息进行推断, 从而可以获得高精度、稠密的深度图.

Multi-view Dense Depth Map Estimation through Match Propagation

The paper proposes a highly accurate multi-view dense depth map estimation algorithm through match propagation. The algorithm is composed of two match propagation processes: a pixel-level propagation process and a region-level propagation process. The former mainly propagates sparse feature matches between views to obtain an initial depth map, the latter, based on multiple initial depth maps and piecewise planar scene assumption, tackles the matching ambiguity problem of some regions (e.g. low texture region) by adopting plane fitting and multiple direction plane sweeping within the framework of minimizing some special energy function. Experiments on standard data set and real-world data set show that our proposed algorithm not only has better adaptability to many factors, e.g. perspective distortion and illumination variance, but also can effectively resolve the depth estimation problem of low texture regions and obtain more accurate and dense depth maps.