# Accurate 3D Body Shape Regression using Metric and Semantic Attributes

**Authors:** Vasileios Choutas, Lea Muller, Chun-Hao P. Huang, Siyu Tang, Dimitrios Tzionas, Michael J. Black  
**arXiv:** [2206.07036](https://arxiv.org/abs/2206.07036)  
**Date:** 2022-06-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2206.07036v1/

## Key points

- The paper addresses the challenge of accurately estimating 3D body shapes from images
- Authors aim to overcome limitations of existing methods in capturing true human shape accurately
- Lack of sufficient data is a main issue in accurate body shape estimation
- Authors utilize internet images of diverse "fashion" models paired with anthropometric measurements and linguistic shape attributes
- Development of SHAPY - a novel neural network that utilizes both anthropometric measurements and linguistic shape attributes for 3D human pose and shape regression from RGB images
- SHAPY outperforms state-of-the-art methods in terms of 3D body shape estimation on existing public benchmarks
- Introduction of new dataset HBW (Human Bodies in the Wild) for further validation, showing significant improvements in 3D body shape estimation compared to existing methods
- Accurate regression of 3D body shapes from images achieved using easily obtainable anthropometric measurements and linguistic shape attributes
- Research outcomes provide valuable insights into advancing computer vision field and have implications for applications requiring precise body shape estimations

## Summary

The paper titled "Accurate 3D Body Shape Regression using Metric and Semantic Attributes" by Vasileios Choutas, Lea Muller, Chun-Hao P. Huang, Siyu Tang, Dimitrios Tzionas, and Michael J. Black addresses the challenge of accurately estimating 3D body shapes from images. The authors aim to overcome the limitations of existing methods in capturing the true human shape accurately. One of the main issues in accurate body shape estimation is the lack of sufficient data. While labeling 2D joints for pose estimation is relatively straightforward for humans, labeling 3D body shapes is a more complex task. To address this problem, the authors utilize two sources of information: internet images of diverse "fashion" models paired with anthropometric measurements and linguistic shape attributes associated with a wide range of 3D body meshes and model images. This approach leads to the development of SHAPY - a novel neural network that utilizes both anthropometric measurements and linguistic shape attributes to regress 3D human pose and shape from RGB images. The performance of SHAPY is evaluated on existing public benchmarks, showcasing its superiority over state-of-the-art methods in terms of 3D body shape estimation. To further validate their approach, the authors introduce a new dataset named HBW (Human Bodies in the Wild), consisting of photos with ground-truth 3D body scans. On this new benchmark dataset, SHAPY demonstrates significant improvements in 3D body shape estimation compared to existing methods. This study represents a groundbreaking demonstration that accurate regression of 3D body shapes from images can be achieved using easily obtainable anthropometric measurements and linguistic shape attributes. The research outcomes presented in this paper provide valuable insights into advancing the field of computer vision and have implications for various applications requiring precise body shape estimations. The model developed by the authors along with their dataset are publicly available at shapy.is.tue.mpg.de for further research and development purposes.