A Feature Aggregation Hourglass Network for Human Pose Estimation
Fan Lü, Yan Zhu Yang · 2021
The fast development of Convolutional Neural Network (CNN) models has greatly contributed to the advancement of human posture estimation. However, hard cases such as invisible keypoints, variable poses and complex background are still great challenges for human pose estimation tasks. To this end, we present a novel cross stage Feature Aggregation Hourglass (FAH) network which aims to improve human pose estimation accuracy. Specifically, we construct a new cascaded residual block for hourglass model and design a multistage hourglass network with cross stage feature aggregation to obtain contextual information from different stages. Moreover, we add transposed convolution layers over our FAH model for upsampling in order to generate heatmaps from deep and low resolution features. Extensive experiments on two common human pose datasets demonstrate the advantages of our FAH model over a broad range of existing human pose estimation methods.