Scale-aware heatmap representation for human pose estimation

doi:10.1016/j.patrec.2021.12.018

Scale-aware heatmap representation for human pose estimation

2022-02-01

10

PRE

AI

DOI：10.1016/j.patrec.2021.12.018

Original

Original request for help

Save

Abstract

En 中文

The performance of multi-person pose estimation is seriously affected by scale variation. Extensive works have been devoted to reducing the effect by modifying convolutional network structure or loss function, but little attention has been paid to the problem in the construction of heatmaps. In this paper, we focus on the scale variation of keypoints within heatmap generation and propose a novel method called scale aware heatmap generator, which constructs a customized heatmap for each type of keypoints based on their relative scales. In addition, we design a weight-redistributed loss function to facilitate the detection of keypoints that are hard to identify. Our approach outperforms the baseline by nearly 2.5% in average precision and performs on par with the state-of-the-art result in bottom-up pose estimation with multi scale testing (69.4% AP) on the COCO test-dev dataset. (c) 2022 Elsevier B.V. All rights reserved.

Keywords:

Human pose estimation

Multi-person

Heatmap representation

Scale-aware heatmap representation for human pose estimation

Abstract

Journal

Organization