UTS-CMU at THUMOS 2015
Zhongwen Xu, Linchao Zhu, Yi Ping Yang, Alexander G. Hauptmann · 2015
This notebook paper describes our solution from UTSCMU team in the THUMOS 2015 action recognition challenge. Our system contains two major components, video representation generated by VLAD encoding from ConvNet features and multi-skip improved Dense Trajectories. In addition, we explore optical flow ConvNet and acoustic features such as MFCC and ASR in our system. We demonstrate that our complete system can achieve state-of-the-art performance in large-scale action recognition tasks.