LiteNetPose: A lightweight neural network for human pose estimation using attention modules

Abstract

This paper presents the usage of attention modules to tackle the challenging problem of the self-occlusion cases in human pose estimation problem. The proposed approach first obtains the relevant features of the human body joints of a set of images using ResNet-50 architecture (just 5.5\% of the 25.6M parameters available are considered) as backbone, which are captured from different views at the same time. Then, a Bone position encoding is proposed to obtain the information about position and orientation of body bone, mainly, those bones whose body joints have more probability to be occluded due to the natural human body pose. These obtained results together with the obtained relevant features of the human body joints using ResNet-50, are used as input to the attention module. Basically, the body joints from a given view are used to enhance poorly estimated joints from another view due to the self-occlusion cases. Experimental results and comparisons with the state-of-the-art approaches on Human3.6m dataset are presented showing improvements in the accuracy of body joints estimations.