Human Pose Estimation Based on the Multistage Learning and the Dense Connection
Weimin Shi, Qiaoning Yang, Juan Chen · 2020
The Human pose estimation is an important and challenging task for the computer vision. Existing human pose estimation networks often improve the accuracy of the pose estimation by feature fusion. However, most of them still suffer from information loss since they only consider high level features. In order to solve this problem, we propose a novel Multistage Learning Network named MS-Net. The key idea behind our approach is that different characteristics of joints may require different levels of feature. MS-Net first predicts joints at different stages of the network in a coarse-to-fine manner. By doing so, both geometric and semantic characteristics of the pose can be learned. Then, to deeper the networks understanding towards the pose, a dense connection is utilized to multiplex multi-level features, which further compensates for the loss of low-level features. We conduct comprehensive experiments on the coco dataset and results show that our model achieves remarkable improvements over state-of-the-art baselines.