The Very Deep Multi-stage Two-stream Convolutional Neural Network for Action Recognition

DOI

10.2991/icmit-16.2016.46How to use a DOI?

Keywords

action recognition; convolutional neural network; multi-stage training

Abstract

In this paper, we consider the very deep multi-stage two-stream convolutional neural network for action recognition in videos. The challenge of action recognition is to capture the appearance and motion information to describe various actions efficiently and to classify different levels of difficult videos correctly. The proposed new deep architecture we name the very deep two-stream convolutional neural network has preferable model capacity and it enables us to obtain appearance and motion information validly from image frames in videos. Besides, with the proposed multi-stage training strategy, multiple classifiers are jointly optimized to process samples at different difficulty levels. Finally, the Dynamic Random Forests classifier is employed to replace Softmax classifier or SVM, achieving a decent classification result. Our architecture is trained and evaluated on the standard video actions benchmarks of UCF-101, and it is competitive with the state of the arts.

Copyright

© 2016, the Authors. Published by Atlantis Press.

Open Access

This is an open access article distributed under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).

Download article (PDF)

TY  - CONF
AU  - Xiuju Gao
AU  - Hanling Zhang
PY  - 2016/04
DA  - 2016/04
TI  - The Very Deep Multi-stage Two-stream Convolutional Neural Network for Action Recognition
BT  - Proceedings of the 2016 3rd International Conference on Mechatronics and Information Technology
PB  - Atlantis Press
SP  - 265
EP  - 269
SN  - 2352-538X
UR  - https://doi.org/10.2991/icmit-16.2016.46
DO  - 10.2991/icmit-16.2016.46
ID  - Gao2016/04
ER  -