POLAR: Posture-level Action Recognition Dataset

摘要

Action category definitions in existing datasets are inconsistent since actions can be labeled at different levels. In this paper, the authors focus on the less studied posture-level actions and introduce a new image dataset, “POsture-Level Action Recognition” (POLAR). The POLAR dataset has a total of 35,324 images and covers 99% of posture-level human actions based on our analysis on the VOC dataset. For action recognition on the POLAR dataset, irrelevant objects or backgrounds could mislead human action predictions. The authors propose the Human-centric Contextual Regions Network (HCRN) that extracts pose related contextual features according to the person in question to address this problem. HCRN outperforms the prior best method in the VOC action classification competition on the POLAR dataset, also reports the new state-of-the-art result, 92.2% with 2.0% relative improvement on the PASCAL VOC action dataset and achieves a promising result on the Stanford 40 dataset.