Almost unsupervised cycle and action detection

Abstract

An event detection method can include encoding a plurality of training video snippets into low dimensional descriptors of the training video snippets in a code space. The low dimensional descriptors of the training video snippets can be decoded into corresponding reconstructed video snippets. One or more parameters of the encoding and decoding can be adjusted based on one or more a loss functions to reduce a reconstruction error between the one or more training video snippets and the corresponding one or more reconstructed video snippets, to reduce a class entropy of the plurality of event classes of the code space, to increase fit of the training video snippet, and/or to increase compactness of the code space. The method can further include encoding one or more labeled video snippets of a plurality of event classes into low dimensional descriptors of the labeled video snippets in the code space. The plurality of event classes can be mapped to class clusters corresponding to the low dimensional descriptors of the labeled video snippets. After training, query video snippets can be encoded into corresponding low dimensional descriptors in the code space. The low dimensional descriptors of the query video snippets can be classified based on their respective proximity to a nearest one of a plurality of class cluster of the code space. An event class of the query video snippet can be determined based on the class cluster classification.