Attention-Enhanced Convolutional Pose Machine for 2D Hand Pose Estimation

Integrating convolution neural networks with attention mechanism is a challenging and promising method for many computer vision tasks. In this study we propose AECPM (Attention-Enhanced Convolutional Pose Machine), an efficient and effective model for 2D hand pose estimation based on CPM by a novel embedding of attention into convolutional pose machines. Specifically, we designed a novel but simple attention-enhanced convolutional (AEConv) module to refine the extracted features from monocular RGB images. At the core of the module is the AEConv blocks and the skip connections between them. Experiments on the CMU Panoptic Dataset show that our method achieves excellent performance.