Abstract

Actionness [3] was introduced to quantify the likelihoodof containing a generic action instance at a specific lo-cation. Accurate and efficient estimation of actionness isimportant in video analysis and may benefit other relevant tasks such as action recognition and action detection.This paper presents a new deep architecture for actionnessestimation, called hybrid fully convolutional network (H-FCN), which is composed of appearance FCN (A-FCN)and motion FCN (M-FCN). These two FCNs leverage thestrong capacity of deep models to estimate actionness mapsfrom the perspectives of static appearance and dynamic motion, respectively. In addition, the fully convolutional na-ture of H-FCN allows it to efficiently process videos with arbitrary sizes. Experiments are conducted on the challenging datasets of Stanford40, UCF Sports, and JHMDB to verify the effectiveness of H-FCN on actionness estimation, which demonstrate that our method achieves superior performance to previous ones. Moreover, we apply the esti-mated actionness maps on action proposal generation and action detection. Our actionness maps advance the current state-of-the-art performance of these tasks substantially.