Exploiting textures for better action recognition in low-quality videos

Human action recognition is an increasingly matured field of study in the recent years, owing to its widespread use in various applications. A number of related research problems, such as feature representations, human pose and body parts detection, and scene/object context, are being actively studied. However, the general problem of video quality—a realistic issue in the face of low-cost surveillance infrastructure and mobile devices, has not been systematically investigated from various aspects. In this paper, we address the problem of action recognition in low-quality videos from a myriad of perspectives: spatial and temporal downsampling, video compression, and the presence of motion blurring and compression artifacts. To increase the resilience of feature representation in these type of videos, we propose to use textural features to complement classical shape and motion features. Extensive results were carried out on low-quality versions of three publicly available datasets: KTH, UCF-YouTube, HMDB. Experimental results and analysis suggest that leveraging textural features can significantly improve action recognition performance under low video quality conditions.