SWBPose: A Lightweight and Efficient Method for Human Pose Estimation
Li Ping Fan, Gaofan Ji · 2023
Human pose estimation is a crucial challenge in the field of computer vision, contributing significantly to diverse domains, such as fall detection, security, and healthcare. While most existing models achieve high accuracy by constructing complex network architectures, they often overlook the issue of their impracticability for deployment on edge devices due to substantial computational requirements. In response to this, we propose SWBPose, a lightweight model tailored for efficient pose estimation. Our research demonstrates that the integration of the Swin Transformer Feature Extract (STFE) module can expedite model convergence. We also introduce the novel Dynamic Shift Max activation function, which has empirically been shown to enhance the model's Average Precision (AP) by 0.2%, without necessitating additional computational resources. We validated our model using the COCO2017 dataset, achieving a mean Average Precision (mAP) of 66.01%, with a frame rate of 96.76 FPS, computational complexity of 0.61 Gflops, and a model size of 4.37M parameters when implemented on a GPU. The experimental results clearly demonstrate that our model offers superior performance in balancing the trade-off between accuracy and latency.