An Unsupervised Ensemble Approach for Emotional Scene Detection from Lifelog Videos - Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing - page 140

An Unsupervised Ensemble Approach for
Emotional Scene Detection from Lifelog Videos

Hiroki Nomiya, Atsushi Morikuni, and Teruhisa Hochin

Abstract. An emotional scene detection method is proposed in order to retrieve
impressive scenes from lifelog videos. The proposed method is based on facial ex-
pression recognition considering that a wide variety of facial expression could be
observed in impressive scenes. Most of conventional facial expression techniques
adopt supervised learning methods. This is a crucial problem because preparing suf-
ficient training data requires considerable human effort due to the diversity of facial
expressions observed in lifelog videos. We thus propose a more efficient emotional
scene detection method using an unsupervised facial expression recognition on the
basis of cluster ensembles. Our approach does not require any training data sets
and is able to detect various emotional scenes. The detection performance of the
proposed method is evaluated through an emotional scene detection experiment.

Keywords: Lifelog, video retrieval, facial expression recognition, clustering, en-
semble learning.

1 Introduction

Lifelog means a person's record of life and has recently attracted attention [1][2].
It can be recorded as various types of data such as texts, images, and videos. We
focus on lifelog videos [3] because anyone can easily record his/her own lifelog
videos due to the improved performance and the minimization of recently developed
video cameras. Although lifelog videos will play an important role in recording and
remembering one's life, they have a serious problem that it is difficult to accurately
and efficiently retrieve useful scenes from the enormous amount of video data. As a
result, valuable lifelog videos often remain unused.