Feature Enhancement and Offset Learning for Multi-Stage Human Pose Estimation
Xuchao Xie, Yi Chen, Yinhao Xu, Teng Fei Gao · 2024
Most of the existing studies have focused on human pose estimation at high resolution; however, research on low-resolution scenes has not yet received extensive attention. The quantization error associated with human pose estimation based on heat map regression increases as the resolution of the detected image decreases, thus seriously affecting the model's detection accuracy. Moreover, the low-resolution image itself loses much of the position information, thus increasing the difficulty of key point localization. Detection at low resolution has important applications in real-world environments and represents an urgent problem to be solved. In this paper, we propose a feature-enhanced offset learning model for human pose estimation based on the high-resolution network HRNet, effectively addressing the problem of quantization error arising from Gaussian heatmap encoding of key point location and addressing the problem of losing detailed information in low-resolution images. The model contains a multi-stage network that gradually obtains lower resolution features while maintaining the existing resolution features and improves the feature extraction capability of the backbone model by fusing different scale features. The extracted information can be enriched with positional information, semantic information, and channel context information by the feature enhancement module, thereby alleviating the problem of missing details due to the low-resolution images. This enhanced information is very important for the subsequent prediction of key points. The offset vector field-based model is used to find out the highest response point position of the heat map and take this position as a rough estimation. Subsequently, it regresses the heat map according to the offset vector to get the offset vector of the corresponding position. Finally, the final position prediction is obtained from the combination of the rough key point position and the offset vector, significantly mitigating the effect of quantization error.