Deep CNN Object Features for Improved Action Recognition in Low Quality Videos

Video based action recognition has been an active area of research in recent years. This is due to the many potential applications such as video surveillance, video content analysis, human computer interaction and video archiving. The ongoing trend of research deals with many complex action recognition problems such as appearance, pose and illumination variations but problem video quality is still considered unexplored. Under low quality conditions, the major challenge is to develop robust feature representation methods that possess discriminative capacity for modeling actions. Many feature representation methods that have been proposed in recent years can be classified into two types: hand-crafted and deeply-learned features. Among various hand-crafted methods proposed in literature, STIP, Cuboid, Hessian, dense sampling, dense trajectory (DT) and improved dense trajectories (IDT) are popular choices. These methods generally rely Email Address: [email protected] Fig.1. Sample low quality videos form UCF-11 ‘compressed’ version (top row) and HMDB51 ‘bad’ and ‘medium’ quality subsets (bottom row).