Enhancing the detection of concepts for visual lifelogs using contexts instead of ontologies

Abstract

Automatic detection of semantic concepts in visual media is typically achieved by an automatic mapping from low-level features to higher level semantics and progress in automatic detection within narrow domains has now reached a satisfactory performance level. In visual lifelogging, part of the quantified-self movement, wearable cameras can automatically record most aspects of daily living. The resulting images have a diversity of everyday concepts which severely degrades the performance of concept detection. In this paper, we present an algorithm based on non-negative matrix refactorization which exploits inherent relationships between everyday concepts in domains where context is more prevalent, such as lifelogging. Results for initial concept detection are factorized and adjusted according to their patterns of appearance, and absence. In comparison to using an ontology to enhance concept detection, we use underlying contextual semantics to improve overall detection performance. Results are demonstrated in experiments to show the efficacy of our algorithm.