Detachable Object Detection with Efficient Model Selection
International Conference on Energy Minimization Methods in Computer Vision and Pattern Recognition (EMMCVPR)
We describe a computationally efficient scheme to perform model selection while simultaneously segmenting a short video stream into an unknown number of detachable objects. Detachable objects are regions of space bounded by surfaces that are surrounded by the medium other than for their region of support, and the region of support changes over time. These include humans walking, vehicles moving, etc. We exploit recent work on occlusion detection to bootstrap an energy minimization approach that is solved with linear programming. The energy integrates both appearance and motion statistics, and can be used to seed layer segmentation approaches that integrate temporal information on long timescales.