# Learning Delicate Local Representations for Multi-Person Pose Estimation

**Authors:** Yuanhao Cai, Zhicheng Wang, Zhengxiong Luo, Binyi Yin, Angang Du, Haoqian Wang, Xiangyu Zhang, Xinyu Zhou, Erjin Zhou, Jian Sun  
**arXiv:** [2003.04030](https://arxiv.org/abs/2003.04030)  
**Date:** 2020-03-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2003.04030v3/

## Key points

- The paper presents a novel method called Residual Steps Network (RSN) for multi-person pose estimation.
- RSN efficiently aggregates features with the same spatial size to obtain delicate local representations that retain rich low-level spatial information and result in precise keypoint localization.
- The authors propose an efficient attention mechanism called Pose Refine Machine (PRM) that makes a trade-off between local and global representations in output features and further refines the keypoint locations.
- The proposed approach achieved state-of-the-art results on both COCO and MPII benchmarks without using extra training data or pretrained models.
- Their single model achieved 78.6 on COCO test-dev and 93.0 on MPII test dataset, while ensembled models achieved 79.2 on COCO test-dev and 77.1 on COCO test-challenge dataset.
- The authors' approach won the first place of COCO Keypoint Challenge 2019.
- The source code for RSN is publicly available for further research at https://github.com/caiyuanhao1998/RSN/.
- Overall, this paper's contribution lies in proposing an innovative method for multi-person pose estimation that achieves high accuracy without relying on external resources such as additional training data or pretrained models.

## Summary

In the field of multi-person pose estimation, this paper presents a novel method called Residual Steps Network (RSN) that efficiently aggregates features with the same spatial size to obtain delicate local representations. These representations retain rich low-level spatial information and result in precise keypoint localization. The authors also observe that output features contribute differently to final performance, which leads them to propose an efficient attention mechanism - Pose Refine Machine (PRM). PRM makes a trade-off between local and global representations in output features and further refines the keypoint locations. The proposed approach achieved state-of-the-art results on both COCO and MPII benchmarks without using extra training data or pretrained models. Their single model achieved 78.6 on COCO test-dev and 93.0 on MPII test dataset, while ensembled models achieved 79.2 on COCO test-dev and 77.1 on COCO test-challenge dataset. The authors' approach won the first place of COCO Keypoint Challenge 2019. The source code for RSN is publicly available for further research at https://github.com/caiyuanhao1998/RSN/. Overall, this paper's contribution lies in proposing an innovative method for multi-person pose estimation that achieves high accuracy without relying on external resources such as additional training data or pretrained models.