Hidden Conditional Random Fields for land-use classification

Undirected probabilistic graphical models such as Markov Random Fields (MRFs) and Conditional Random Fields (CRFs) are being increasingly used to model problems having a structured domain and to enable probabilistic inferences such as answering queries about the variables of interest, e.g., inferring classification labels of pixel patches or images. We investigate Multiple-Instance learning approach based on Hidden Conditional Random Fields for land-use classification using weakly labeled aerial images. The performance is evaluated using publicly available dataset that contains aerial imagery belonging to 21 land-use categories.