# LCR-Net++: Multi-person 2D and 3D Pose Detection in Natural Images

**Authors:** Gregory Rogez, Philippe Weinzaepfel, Cordelia Schmid  
**arXiv:** [1803.00455](https://arxiv.org/abs/1803.00455)  
**Date:** 2018-03-01  
**Summary page:** https://www.summarizepaper.com/en/paper/1803.00455v1/

## Key points

- The paper proposes an end-to-end architecture for joint 2D and 3D human pose estimation in natural images.
- The approach generates and scores a number of pose proposals per image, allowing for the prediction of poses of multiple people simultaneously without requiring approximate localization.
- The LCR-Net++ architecture contains three main components: pose proposal generator, classifier, and regressor, all trained jointly.
- Final pose estimation is obtained by integrating over neighboring pose hypotheses to improve upon non-maximum suppression algorithm.
- Approach recovers full-body 2D and 3D poses accurately even when persons are partially occluded or truncated by the image boundary.
- Outperforms state-of-the-art methods in 3D pose estimation on Human3.6M dataset and shows promising results on real images for both single and multi-person subsets of MPII 2D pose benchmark.
- Improvements over previous work include better handling of occlusions through improved data augmentation techniques, incorporating temporal information from video sequences to improve accuracy further, and introducing new evaluation metrics to better assess performance across different datasets.

## Summary

In their paper titled "LCR-Net++: Multi-person 2D and 3D Pose Detection in Natural Images," Gregory Rogez, Philippe Weinzaepfel, and Cordelia Schmid propose an end-to-end architecture for joint 2D and 3D human pose estimation in natural images. The key to their approach is the generation and scoring of a number of pose proposals per image, which allows them to predict 2D and 3D poses of multiple people simultaneously without requiring an approximate localization of the humans for initialization. Their Localization-Classification-Regression architecture, named LCR-Net++, contains three main components: (1) the pose proposal generator that suggests candidate poses at different locations in the image; (2) a classifier that scores the different pose proposals; and (3) a regressor that refines pose proposals both in 2D and 3D. All three stages share the convolutional feature layers and are trained jointly. The final pose estimation is obtained by integrating over neighboring pose hypotheses, which improves upon a standard non-maximum suppression algorithm. Their method recovers full-body 2D and 3D poses accurately even when persons are partially occluded or truncated by the image boundary. Their approach significantly outperforms state-of-the-art methods in 3D pose estimation on Human3.6M, a controlled environment. Moreover, it shows promising results on real images for both single and multi-person subsets of the MPII 2D pose benchmark as well as satisfying 3D pose results even for multi-person images. The authors also introduce several improvements over their previous work on LCR-Net such as better handling of occlusions through improved data augmentation techniques, incorporating temporal information from video sequences to improve accuracy further as well as introducing new evaluation metrics to better assess performance across different datasets. Overall, this paper presents a significant contribution towards improving the accuracy of joint 2D and 3D human pose estimation in natural images with promising results on both controlled and real world datasets.