基于LLOM的单目图像深度图估计算法

针对计算机视觉理解单目图像立体结构的问题,进行了单目图像深度估计算法的研究。提出了一种基于监督学习方法的室外单目图像深度估计算法,其采用语义标注信息指导深度估计过程,融合绝对深度特征、相对深度特征以及位置特征作为深度特征向量,采用LLOM学习深度特征向量与深度值之间的关系。实验结果显示,该算法对路面、草地以及建筑物类等深度渐进变化的图像块,可获得较满意的深度估计结果。本算法为单目图像深度估计开辟了一个全新的有效途径。

Depth estimation algorithm from monocular image based on LLOM

The key work of this paper was to estimate the scene depth for a single monocular image based on machine learning algorithm. Under the direction of semantic labels, an improved ASSOM algorithm named LLOMlocally liear online mapping to depth estimation first time to learn the manifold of high dimension feature vectors, including absolute depth feature, relative depth feature and position feature. The experiments show that proposed method can obtain an acceptable result, especially for those blocks with gradual change of depth. This algorithm provides a new possibility to estimate depth from a single image.