Simple Pose Network with Skip-Connections for Single Human Pose Estimation

Recently, following the success of deep convolutional neural networks, human pose estimation problem has been largely improved. This paper introduces an improved version of the Simple Pose network for single human pose estimation. It adds the skip-connections between the same-resolution layers of the backbone and up-sampling stream to fuse low-level and high-level features. To make the depth of features from low-level and high-level are same, this paper uses \(1\,\times \,1\) convolutional layer. The experiments show that this naive technique makes the new networks better over 1% mAP scores with just a small increment in model size.