This paper proposes a system that relates objects\udin an image using occlusion cues and arranges them according\udto depth. The system does not rely on a priori knowledge of\udthe scene structure and focuses on detecting special points,\udsuch as T-junctions and highly convex contours, to infer the\uddepth relationships between objects in the scene. The system\udmakes extensive use of the binary partition tree as hierarchical\udregion-based image representation jointly with a new approach\udfor candidate T-junction estimation. Since some regions may\udnot involve T-junctions, occlusion is also detected by examining\udconvex shapes on region boundaries. Combining T-junctions and\udconvexity leads to a system which only relies on low level depth\udcues and does not rely on semantic information. However, it\udshows a similar or better performance with the state-of-the-art\udwhile not assuming any type of scene.\udAs an extension of the automatic depth ordering system, a\udsemi-automatic approach is also proposed. If the user provides\udthe depth order for a subset of regions in the image, the system\udis able to easily integrate this user information to the final\uddepth order for the complete image. For some applications, user\udinteraction can naturally be integrated, improving the quality of\udthe automatically generated depth map.
展开▼