Refinement of Ontology-Constrained Human Pose Classification

Abstract

In this paper, we propose an image classification method that recognizes several poses of idol photographs. The proposed method takes unannotated idol photos as input, and classifies them according to their poses based on spatial layouts of the idol in the photos. Our method has two phases, the first one is to estimate the spatial layout of ten body parts (head, torso, upper and lower arms and legs) using Eichner's Stickman Pose Estimation. The second one is to classify the poses of the idols using Bayesian Network classifiers. In order to improve accuracy of the classification, we introduce Pose Guide Ontology (PGO). PGO contains useful background knowledge, such as semantic hierarchies and constraints related to the positional relationship between the body parts. The location information of body parts is amended by PGO. We also propose iterative procedures for making further refinements of PGO. Finally, we evaluated our method on a dataset consisting of 400 images in 8 poses, and the final results indicated that F-measure of the classification has become 15% higher than non-amended results.