Object Priors for Classifying and Localizing Unseen Actions

Abstract This work strives for the classification and localization of human actions in videos, without need any labeled video training examples. Where existing relies on transferring global attribute or object information from seen to unseen action we seek classify spatio-temporally localize videos image-based only. We propose three spatial priors, which encode local person detectors along with their relations. On top introduce semantic extend matching through word embeddings simple functions that tackle ambiguity, discrimination, naming. A embedding combines priors. It enables us a new retrieval task retrieves tubes collections based user-specified objects, relations, size. Experimental evaluation five datasets shows importance priors actions. find persons objects have preferred relations benefit localization, while using multiple languages filtering directly improves matching, leading state-of-the-art results both localization.