Attention Refined Network for Human Pose Estimation

Recently, multi-scale feature fusion has been considered as one of the most important issues in designing convolutional neural networks (CNNs). However, most existing methods directly add the corresponding layers together without considering the semantic gaps between them, which may lead to inadequately feature fusion results. In this paper, we propose an attention refined network (HR-ARNet) to enhance multi-scale feature fusion for human pose estimation. The HR-ARNet employs channel and spatial attention mechanisms to reinforce important features and suppress unnecessary ones. To tackle the problem of inconsistent among keypoints, we utilize self-attention strategy to model long-range keypoints dependencies. We also propose to use the focus loss, which modifies the commonly used square error loss function to let it mainly focus on top K ‘hard’ keypoints during training. Focus loss selects ‘hard’ keypoints based on the training loss and only backpropagates the gradients from the selected keypoints. Experiments on human pose estimation benchmark, MPII Human Pose Dataset and COCO Keypoint Dataset, show that our method can boost the performance of state-of-the-art human pose estimation networks including HRNet (high-resolution net) (Sun et al., Proceedings of the IEEE conference on computer vision and pattern recognition, 2019). The code and models are available at: http://github/tongjiangwei/ARNet.