Supervised model-based depth map creation of humans in video

The success of 3D television depends on the availability of movies, series and television shows in 3D. Semi-automated conversion of existing 2D video to 3D video is an approach where a depth map is produced for each frame in the video. This is done in a semi-automated manner such that there is control on the quality of the 3D output. The 3D quality is determined both by the amount of user interaction and by the cleverness of the used algorithms. Since cost is important for any technology, we constantly wish to improve the quality while keeping required amount of operator interaction limited. This tradeoff is important when designing new algorithms for offline semi-automated 2D to 3D video conversion. One of the methods to create a depth map from a 2D video image is to segment out the objects in the image and assign manually or extract automatically the depth information of the objects. In this case the most interesting group of segmented objects is humans. When we are looking at an image, we are more interested in the humans in the picture than in other objects. Furthermore, we are more sensitive to the details in the human body and face. Therefore, defining a method to create depth information particularly for the humans in the 2D video material would provide a better perception on 3D displays. In this paper we describe a system to interactively create a depth map for pictures that contain humans