eduzhai > Applied Sciences > Engineering >

TinyVIRAT Low-resolution Video Action Recognition

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 9 pages

Abstract: The existing research in action recognition is mostly focused on high-qualityvideos where the action is distinctly visible. In real-world surveillanceenvironments, the actions in videos are captured at a wide range ofresolutions. Most activities occur at a distance with a small resolution andrecognizing such activities is a challenging problem. In this work, we focus onrecognizing tiny actions in videos. We introduce a benchmark dataset,TinyVIRAT, which contains natural low-resolution activities. The actions inTinyVIRAT videos have multiple labels and they are extracted from surveillancevideos which makes them realistic and more challenging. We propose a novelmethod for recognizing tiny actions in videos which utilizes a progressivegenerative approach to improve the quality of low-resolution actions. Theproposed method also consists of a weakly trained attention mechanism whichhelps in focusing on the activity regions in the video. We perform extensiveexperiments to benchmark the proposed TinyVIRAT dataset and observe that theproposed method significantly improves the action recognition performance overbaselines. We also evaluate the proposed approach on synthetically resizedaction recognition datasets and achieve state-of-the-art results when comparedwith existing methods. The dataset and code is publicly available atthis https URL.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×