Discriminative EdgeBoxes for Action Recognition

Due to the huge amount of online videos uploaded and viewed every day, there is an emerging need nowadays for action recognition techniques. These techniques need to consider the large variations in camera motion, viewpoint, cluttered background, etc. Moreover, they need to be unsupervised or weakly supervised to can absorb such amount of different actions. The goal of this paper is to introduce a new unsupervised technique for mining mid-level discriminative patches from videos. These patches are the most representative parts that can describe an action. To achieve this goal, we generalize a technique borrowed from 2D images to generate bounding boxes with high motion and appearance saliencies then apply iterative clustering/classification procedure on generated boxes. Then, we calculate discriminative score for each box and finally select top ranked boxes to train Exemplar-SVM on low-level features extracted inside selected boxes. We evaluate our approach on a challenging dataset namely YouTube and the experimental results demonstrate the effectiveness of our approach to achieve better average recognition accuracy than the state-of-the-art techniques.