Adaptive 3-D scene construction from single image using extended object placement relation

Object placement relation (OPR) is the information set representing relationship between set of image components. It can be applied for constructing various types of image scenes quickly and efficiently ranging from scenes with multi-level of the ground, natural scene, building scene to outdoor scene. OPR can also be used to predict depth of the scene without reference ground whereas other methods cannot. However, the result of using previous OPR shows some error when applying to indoor scenes with ceiling height because OPR lacks such relationship. Moreover, there still exists some distortion and unsmoothness between image components due to the local mesh computation of each image component. In this research the OPR is extended by adding more constraints to each relation, while the triangulation algorithm is adapted to preserve the global smoothness of the output 3-D scene. Our experimental result shows that, with the similar time complexity, the adaptive OPR can improve the quality of 3-D scenes with smoother surface than when using the previous OPR.