Temporally consistent tracking and segmentation of people in video

This will research methods for live (video-rate)temporally consistent segmentation and tracking of non-rigid objects such as people in video of general scenes in order to enable seamless real-time scene augmentation (e.g. insertion of graphical elements for advertising, appearance editing). The project will primarily address the decomposition of a scene into its constituting layers and the inference of a dense (surface-point level) representation of their motion. Other complementary aspects that may be explored as part of the research include extraction of appearance/shading to further improve the representation and the realism of the generated content. Target applications of this research will include sports TV footage, TV broadcast footage and YouTube video content.