基于单幅图像的深度非连续性估计

Depth discontinuity estimation based on single image

如何准确保持深度的非连续是以视觉方法进行三维重建中的重要问题。针对此类问题,提出了一种基于单幅图像进行几何非连续性估计的全局优化方法。将其作为一个统计学习问题,构建由一系列彩色图像及其对应深度图组成的训练库,通过在图像和深度图中分别选取合适的图像特征向量以及深度连续性度量,训练联合高斯马尔可夫随机场概率模型,使得几何非连续估计过程能够有效结合图像的上下文信息来进行全局推理。实验结果证明了算法的有效性。

How to preserve depth discontinuity is an important issue needed to be considered in most work on visual 3D reconstruction.This paper attempted to extract this information from a single image,and formulate the perception of depth discontinuity as a statistical learning problem.A training set was created,which is composed of a series of real-color images and their corresponding depthmaps.Then,by extracting appropriate image feature and measuring corresponding depth discontinuity,a joint Gaussian Markov random field was trained to model the conditional distribution of depth discontinuity and gave the various visual cues.The experimental results show that the presented algorithm is able to recover fairly accurate depth discontinuity maps.