Global for Coarse and Part for Fine: A Hierarchical Action Recognition Framework

Abstract

Action recognition is one significant yet challenging task in computer vision. Recent methods mainly model an end-to-end one-stage non-deep or deep learning networks to distinguish different action categories. In this paper we introduce one novel hierarchical action classification framework: Unlike existing one-stage recognition models, the proposed work improves the recognition accuracy by: 1) developing a hierarchical coarse-to-fine action classification framework by dividing the recognition processing into two stages: coarse- grained classification and fine-grained classification, and 2) representing actions in different stages with different granularity features representation: global features are utilized for coarse classifiers while more body parts patterns for fine-grained classifiers are aggregated. Experiments on two widely-tested benchmark datasets show that our method can achieve state-of-the-art or competitive performance compared with existing results using one-stage models, with advantages regarding the recognition accuracy and robustness.