Multi-Person Pose Estimation via Learning Feature Integration

Abstract

Human pose estimation is a fundamental but challenging task in computer vision. Although batch normalization is widely used for deep learning, feature extraction in deep convolutional neural networks (DCNN) is still not well explored. In this work, we propose a pose encoding module (PEM) to enhance the learning ability and generalization ability of feature extraction. Given input images, PEM integration instance normalization and batch normalization, combining them in an appropriate way to learn to capture and eliminate appearance changes while maintaining the distinction between learning features can be seen as an integration and adjustment of global information. In addition, we use a simple and efficient up-sampling strategy to recover high-resolution representations for predicting more accurate human keypoint heatmaps, which has achieved better performance than the average network in recent missions. We studied our approach on a standard benchmark for human pose estimation.