A Comprehensive Study of Feature Representations for Semantic Concept Detection

Abstract

In this paper, we study the performance of global features and local features for the semantic concept detection problem, which is a crucial task for video indexing and retrieval applications. We have performed a comprehensive evaluation of these features on TRECVID datasets that are used in the concept detection benchmark from 2005 to 2009. The experimental results show that with appropriate choice of parameters, global features such as local binary patterns and edge orientation histogram can achieve reasonable performance compared with local features using BoW model while requiring fewer number of parameters to be tuned and lower computational cost. Furthermore, we also investigate on how to design a compact concept detection system that can balance between computational cost and accuracy.