Simplification and Optimization of Visual Tracking Convolutional Neural Networks Parameters

DOI

10.2991/icmmita-16.2016.46How to use a DOI?

Keywords

Visual tracking; convolutional neural network; dynamic compression.

Abstract

Convolutional Neural Network (CNN), which has achieved great success in traditional computer vision tasks such as image classification, object detection, scene segmentation, etc. has also gained great success in fields of visual tracking. However, the algorithm based on CNN is both computation and memory intensive, which makes it hard to deploy it on various systems, especially for embedded system. In this paper, we choose the champion of VOT 2015 challenge which is multi-domain convolutional neural network for visual tracking (MDNet) as our algorithm prototype. We eliminate the zero value and even more non-zero intermediate values in the process of algorithm to reduce both computation and memory resources. While the threshold of non-zero values is not easily determined, we experiment some thresholds and give the result of compression ratio and tracking accuracy to find the relationship between thresholds and tracking accuracy. Finally we achieve some relationship such as nearly 50% compression ratio with 2.3% area-under-the-curve (AUC) loss and analyze the reason of the case in which algorithm is failed.

Copyright

© 2017, 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  - Zhiyong Qin
AU  - Lixin Yu
PY  - 2017/01
DA  - 2017/01
TI  - Simplification and Optimization of Visual Tracking Convolutional Neural Networks Parameters
BT  - Proceedings of the 2016 4th International Conference on Machinery, Materials and Information Technology Applications
PB  - Atlantis Press
SP  - 253
EP  - 259
SN  - 2352-538X
UR  - https://doi.org/10.2991/icmmita-16.2016.46
DO  - 10.2991/icmmita-16.2016.46
ID  - Qin2017/01
ER  -