Two-Stream convolutional nets for action recognition in untrimmed video

Abstract

We extend the two-stream convolutional net architecture developed by Simonyan for action recognition in untrimmed video clips. The main challenges of this project are first replicating the results of Simonyan et al, and then extending the pipeline to apply it to much longer video clips in which no actions of interest are taking place most of the time. We explore aspects of the performance of the two-stream model on UCF101 to elucidate the current barriers to better performance in both UCF101 and untrimmed video. We also explore whether or not training on the background videos from the Thumos Challenge dataset (in which no actions of interest occur in the video) improves action recognition.

6 Figures and Tables

Cite this paper

@inproceedings{Jung2015TwoStreamCN, title={Two-Stream convolutional nets for action recognition in untrimmed video}, author={Kenneth Jung and Song Han}, year={2015} }