Structural Edge Learning for 3-d Reconstruction from a Single Still Image

Learning 3-D scene structure from a single still image has become a hot research topic recently, during which edges played a very important role as they provided critical information about the structures. While not all the intensity edges are useful, existing 3-D reconstruction methods suffered heavily when not differentiating structural edges with non-structural ones. In this report, we consider learning structural edges rather than edges from intensity values of the image. Through supervised learning, the learnt edges as shown in this report carries more stuctural information and less noise than intensity edges. The comparison of two kinds of edges are also shown in the report.