Keeping count: Leveraging temporal context to count heavily overlapping objects

When tracking and segmenting multiple objects under heavy occlusion, a large class of algorithms can greatly benefit from a preprocessing that reliably assesses the number of individuals in each cluster. This is a difficult task when relying on local information only, due to scarcity of training examples and lack of strongly predictive features. In this paper, we develop a deterministic graphical model to address the problem of counting the number of objects in each foreground region as global inference across the entire video sequence. We show that global inference improves over local predictions, and is able to produce an accurate and coherent output within an useful runtime.