# ICON: Implicit Clothed humans Obtained from Normals

**Authors:** Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, Michael J. Black  
**arXiv:** [2112.09127](https://arxiv.org/abs/2112.09127)  
**Date:** 2021-12-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2112.09127v1/

## Key points

- Authors address limitations of current methods for learning 3D clothed avatars
- Proposed method uses implicit functions to capture intricate details like hair and clothing
- Introduces ICON framework leveraging local features for diverse human poses
- Utilizes multiple frames and SCANimate for creating animatable avatars with superior performance
- Evaluation on AGORA and CAPE datasets shows robustness to out-of-distribution samples
- Represents significant advancement in 3D clothed human reconstruction from real-world images

## Summary

In the paper "ICON: Implicit Clothed humans Obtained from Normals," authors Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black address the limitations of current methods for learning realistic and animatable 3D clothed avatars. The proposed method utilizes implicit functions to capture intricate details such as hair and clothing in the avatar reconstruction process. To overcome challenges with diverse human poses, the authors introduce ICON, a novel framework that leverages local features instead of global ones. By reconstructing multiple frames of a subject in different poses and utilizing SCANimate, the method can create an animatable avatar with superior performance compared to state-of-the-art techniques. Evaluation on AGORA and CAPE datasets demonstrates ICON's robustness to out-of-distribution samples like in-the-wild poses/images and out-of-frame cropping. This approach represents a significant advancement towards achieving robust 3D clothed human reconstruction from real-world images.