Compact video representation for video event retrieval and recognition

摘要

Comprehensive, compact, and discriminative representations of videos can be obtained using a counting grid representation of the video and aggregating features associated with active locations of the counting grid to obtain a feature representation of the video. The feature representation can be used for video retrieval and/or recognition. In some examples, the techniques may include conducting normalization and dimension reduction on the aggregated features to obtain a further compact and discriminative feature representation. In some examples, the counting grid representation of the video is generated using a pre-trained counting grid model in order to provide spatially consistent feature representations of the videos.