2D Human Pose Estimation with Residual Upsampling and Spatial Feature Enhancement
Shanbin Li, Xishan Zheng · 2024
Computer vision, which has been extensively used in intelligent monitoring, self-driving, medical assistance, sports analysis, and other fields, includes human pose estimation as a key component. Human pose estimation algorithms have been appearing nonstop in recent years, and as they get more complex and structured, it becomes harder to analyze and actually implement the algorithms. SBL network is surprisingly efficient for such a simple design. This study makes several modifications based on SBL to further enhance its performance. In order to fuse rich features from the feature extraction stage to the upsampling phase, we first introduce the residual upsampling module (RUSM). This guarantees that more detailed, high-resolution information can be utilized to upsampling features, significantly increasing the accuracy of joint recognition in the human body. Additionally, we present the spatial feature enhancement module (SFEM), which utilizes the spatial attention mechanism with the residual structure. It is employed to boost spatial features of the upsampling phase and further increase the detection precision of the network. Lastly, we conduct numerous experiments on the LSP dataset, which amply illustrates the enhanced functionality of the extra modules and the better performance of our suggested network.