Boosted Exemplar Learning for human action recognition

Human action recognition has been an active research topic in computer vision. How to model all kinds of actions, varying with time resolution, visual appearance, etc., is quite a challenging task for recognition. In this paper, we propose a Boosted Exemplar Learning (BEL) approach to recognize various actions in a weakly supervised manner, i.e., only video-based labels are provided but frame-based ones are not. First, for a given action, each video is described as a set of similarities between its frames and some candidate ones (called as exemplars), which are selected from training videos belonging to the action. Instead of simply using a heuristic distance measure, the similarities are decided by the exemplar-based classifiers through the Multiple Instance Learning (MIL), in which a positive (or negative) video is deemed as a positive (or negative) bag and those similar frames to the given exemplar in Euclidean Space as instances. Second, we formulate the selection of the most discriminative exemplars into a boosted feature selection framework and simultaneously obtain a video-based action detector in the boosted learning process. Experimental results on two publicly available challenging datasets: the KTH dataset and Weizmann dataset demonstrate the validity and effectiveness of the proposed approach.