A lightweight network for human keypoint detection based on hybrid attention
Youtao Luo, Xiaoming Gao · 2024
In response to the challenges of high parameter count, computational complexity, and long inference time in current human pose estimation network models, we have combined the characteristics of self-attention mechanism, convolutional networks, and channel attention to propose a hybrid attention module called CSA-Block. This module allows for flexible selection of different attention mechanisms based on the depth and width of the network. By doing so, we not only ensure the detection performance of the network but also reduce the computational cost and improve the inference speed. Furthermore, based on the CSA-Block module, we have designed a multi-resolution and multi-branch architecture and applied it to pose estimation tasks. Our experiments on the MSCOCO and MPII datasets demonstrate that under the condition of minimal sacrifice in accuracy, our approach significantly reduces the parameter count of the network while greatly accelerating the inference time.