Optimized S2E Attention Block based Convolutional Network for Human Pose Estimation

Ya‐Pei Feng

doi:10.1109/ACCESS.2022.3216470

Return

Optimized S2E Attention Block based Convolutional Network for Human Pose Estimation

2022-01-01

Ya‐Pei Feng

Penghui Liu

Zhe‐Ming Lu *

DOI：10.1109/ACCESS.2022.3216470

Original

Original request for help

Save

Abstract

En 中文

Human pose estimation is a popular research area due to its wide range of application scenarios. The general works, on the other hand, concentrate on how to enhance the network's width, depth, and resolution, which results in a sizable number of parameters that hinder practical implementation on real-time and resource-constrained devices. Furthermore, since networks are severely constrained by the unequal feature distribution, it is challenging to extract deep features. We propose an S2E-based attention module that is lightweight, easily scalable, and aims to achieve a balance between reignition accuracy and speed while using fewer computational resources. The optimized S2E attention model consists of two layers of compression modules and one layer of motivation modules. We apply this attention block to the classical ResNet-101 and HRNet network backbones to build our S2E Based S2E-ResNet-101 and S2E-HRNet structure. Comparative studies on the COCO dataset and the MPII dataset show that the S2E module consumes very few computational resources but shows significant improvement in prediction accuracy, achieving a better speed/accuracy tradeoff and being more practical than other state-of-the-art methods. Moreover, the visual output of the qualitative comparison experiments in chaotic pose recognition further demonstrates our model's capacity to concentrate on a significantly more detailed area and prevent erroneous recognition brought on by posture crossover and occlusion. Overall, it can be seen that the S2E module is a simple but effective and easily scalable attention module, which is of tremendous practical value to the field of pose recognition.

Keywords:

Feature extraction

Convolutional neural networks

Computational modeling

Pose estimation

Correlation

Neural networks

Human factors

S2E model

attentional block

pose estimation