# Dynamic Multi-Person Mesh Recovery From Uncalibrated Multi-View Cameras

**Authors:** Buzhen Huang, Yuan Shu, Tianshu Zhang, Yangang Wang  
**arXiv:** [2110.10355](https://arxiv.org/abs/2110.10355)  
**Date:** 2021-10-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2110.10355v1/

## Key points

- Dynamic multi-person mesh recovery from uncalibrated multi-view cameras is a challenging task in the field of 3D vision.
- Few works have specifically focused on multi-person motion capture from uncalibrated cameras.
- The problem presents two main challenges: inter-person interactions and occlusions introduce ambiguities for camera calibration and motion capture, and the lack of dense correspondences makes it difficult to constrain sparse camera geometries in a dynamic multi-person scene.
- The authors propose a novel approach that incorporates motion prior knowledge into the simultaneous optimization of extrinsic camera parameters and human meshes using slightly noisy inputs.
- The key idea is to leverage physics-geometry consistency to reduce low and high-frequency noises in the detected human semantics.
- By introducing a latent motion prior, accurate camera parameters and coherent human motions can be obtained through one-stage optimization.
- Experimental results demonstrate the effectiveness of the proposed method, showing successful recovery of accurate camera parameters and human motions even with inter-person interactions and occlusions present.
- The authors plan to make their codes publicly available for further research and development.

## Summary

Dynamic multi-person mesh recovery from uncalibrated multi-view cameras is a challenging task in the field of 3D vision. While there has been significant research on this topic, few works have specifically focused on multi-person motion capture from uncalibrated cameras. This problem presents two main challenges: inter-person interactions and occlusions introduce inherent ambiguities for both camera calibration and motion capture; and the lack of dense correspondences makes it difficult to constrain sparse camera geometries in a dynamic multi-person scene. To address these challenges, the authors propose a novel approach that incorporates motion prior knowledge into the simultaneous optimization of extrinsic camera parameters and human meshes using slightly noisy inputs. The key idea is to leverage physics-geometry consistency to reduce low and high-frequency noises in the detected human semantics. By introducing a latent motion prior, accurate camera parameters and coherent human motions can be obtained through one-stage optimization. The experimental results demonstrate the effectiveness of the proposed method, showing successful recovery of accurate camera parameters and human motions even with inter-person interactions and occlusions present. The authors plan to make their codes publicly available for further research and development. In conclusion, this paper presents an effective approach for dynamic multi-person mesh recovery from uncalibrated multi-view cameras that contributes to advancing the field of 3D vision by addressing the challenges associated with multi-person motion capture from uncalibrated cameras.