Human pose estimation via improved ResNet50

Abstract

This paper provides a method to predict 2D human pose in an image based on deep model ResNet-50. Human pose estimation is formulated as a regression problem towards body joints through top-down methods. First, we detect the position of humans in holistic image. Then, we take advantages of multi-stages cascade of ResNet-50 to reason about human body joints position. Our approach on challenging the FLIC datasets with large pose variation outperforms the state-of-the-art methods on these benchmarks.