Multiple View Object Cosegmentation Using Appearance and Stereo Cues

摘要: We present an automatic approach to segment object in calibrated images acquired from multiple viewpoints. Our system starts with a new piecewise planar layer-based stereo algorithm that estimates dense depth map consists of set 3D surfaces. The is formulated using energy minimization framework combines and appearance cues, where for each surface, model learnt unsupervised approach. By treating the surfaces as structural elements scene reasoning about their visibility views, we image independently. Finally, these segmentations are refined by probabilistically fusing information across views. demonstrate our can challenging objects complex shapes topologies, which may have thin structures non-Lambertian It also handle scenarios background color distributions overlap significantly.