Abstract:Human pose estimation, a vital pursuit in the realm of computer vision, aims to predict the spatial coordinates of key points within images. Despite the advancements achieved by employing a Convolution Neural Network (CNN), this task still faces considerable challenges, especially in handling occlusion and overfitting issues. This paper introduces a new human pose estimation network designed to address the challenges posed by occluded and blurred images. It features a multi-scale spatial attention mechanism that zeroes in on the human body, significantly improving feature extraction for complex images. Moreover, this versatile attention module is compatible with a wide range of convolutional neural network-based pose estimation frameworks, unlike other mechanisms restricted to particular networks. Addressing the overfitting issue in human pose estimation models, this paper introduces an adversarial network-based data augmentation technique. A generator specifically tailored for pose estimation is adversarially trained to produce optimal augmentation samples, thereby reducing model overfitting. Experimental validation confirms that this augmentation method notably enhances the prediction accuracy of the pose estimation model without incurring extra computational costs. In addition, this paper introduces a streamlined Feature Pyramid Network (FPN) that enables shallow networks to assimilate extensive-scale data, addressing the issue of excessive model size. The experimental validation on the benchmark datasets MPII and MSCOCO demonstrates the efficacy of this integrated approach, showcasing significant improvements in the accuracy and the overall performance of human pose estimation and surpassing the existing methodologies. This approach effectively enhances the performance of the baseline model, achieving the best accuracy of 92.2% and 80.4% on the MPII and MSCOCO, respectively.

Human Pose Estimation Based on Lightweight Multi-Scale Coordinate Attention

X-HRNet: Towards Lightweight Human Pose Estimation with Spatially Unidimensional Self-Attention

Lightweight Human Pose Estimation Based on Adaptive Feature Sensing

Simple and Lightweight Human Pose Estimation

Research on Multi-level Attention-based Human Pose Estimation

Lightweight Non-local High-Resolution Networks for Human Pose Estimation.

Lightweight high-resolution network based on adaptive cross-dimensional weighting for human pose estimation

A Multi-Level Network for Human Pose Estimation

Enhancement and Optimisation of Human Pose Estimation with Multi-Scale Spatial Attention and Adversarial Data Augmentation

Efficient and Lightweight Human Pose Estimation Algorithm

Human Pose Estimation Based on Efficient and Lightweight High-Resolution Network (EL-HRNet)

MSPENet: Multi-Scale Adaptive Fusion and Position Enhancement Network for Human Pose Estimation

Multi-Scale Adaptive Structure Network For Human Pose Estimation From Color Images

A lightweight pose estimation network with multi-scale receptive field

Human Pose Estimation Based on Feature Enhancement and Multi-Scale Feature Fusion

Lightweight Human Pose Estimation Using Loss Weighted by Target Heatmap.

A Compact and Powerful Single-Stage Network for Multi-Person Pose Estimation

AdaptivePose++: A Powerful Single-Stage Network for Multi-Person Pose Regression

A Deconvolutional Bottom-up Deep Network for Multi-Person Pose Estimation.

Ghost attentional down net: An effective lightweight top-down network for human pose estimation

Multi-Scale Supervised Network for Human Pose Estimation