# Norface: Improving Facial Expression Analysis by Identity Normalization

**Authors:** Hanwei Liu, Rudong An, Zhimeng Zhang, Bowen Ma, Wei Zhang, Yan Song, Yujing Hu, Wei Chen, Yu Ding  
**arXiv:** [2407.15617](https://arxiv.org/abs/2407.15617)  
**Date:** 2024-07-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.15617v1/

## Key points

- Facial Expression Analysis is a complex task with task-irrelevant noise factors such as identity, head pose, and background.
- The Norface framework, introduced by authors Hanwei Liu, Rudong An, Zhimeng Zhang, Bowen Ma, Wei Zhang, Yan Song, Yujing Hu, Wei Chen and Yu Ding in this paper, addresses the challenge by using a two-part system consisting of a normalization network and a classification network.
- The normalization network removes task-irrelevant noise by ensuring consistency in facial expressions through normalizing all original images to a common identity with consistent pose and background.
- The classification network incorporates a Mixture of Experts to refine the latent representation by handling input facial representations and outputting multiple labels for Action Units (AUs) or emotions.
- Extensive experiments validate the effectiveness of the Norface framework in various facial expression analysis tasks such as AU detection, AU intensity estimation, FER tasks, and cross-dataset tasks.
- The proposed method has shown superior performance compared to existing state-of-the-art methods.
- Interested individuals can access the normalized datasets and code related to this research at https://norface-fea.github.io/.
- This paper has been accepted by ECCV2024 and contributes towards improving Facial Expression Analysis through innovative techniques like identity normalization.

## Summary

Facial Expression Analysis is a complex task that involves various task-irrelevant noise factors such as identity, head pose, and background. To address this challenge, a novel framework called Norface has been introduced by authors Hanwei Liu, Rudong An, Zhimeng Zhang, Bowen Ma, Wei Zhang, Yan Song, Yujing Hu, Wei Chen and Yu Ding in this paper. The Norface framework is designed to tackle both Action Unit (AU) analysis and Facial Emotion Recognition (FER) tasks through a two-part system consisting of a normalization network and a classification network. The normalization network plays a crucial role in removing task-irrelevant noise by ensuring consistency in facial expressions. It achieves this by normalizing all original images to a common identity with consistent pose and background. This step enables the classification network to effectively extract useful expression information from the normalized images. Additionally, the classification network incorporates a Mixture of Experts to refine the latent representation by handling input facial representations and outputting multiple labels for AUs or emotions. Extensive experiments have been conducted to validate the effectiveness of the Norface framework with insights into identity normalization. The proposed method has shown superior performance compared to existing state-of-the-art methods in various facial expression analysis tasks such as AU detection, AU intensity estimation, FER tasks, and cross-dataset tasks. For those interested in accessing the normalized datasets and code related to this research, they are encouraged to visit https://norface-fea.github.io/. This paper has been accepted by ECCV2024 and provides valuable contributions towards improving Facial Expression Analysis through innovative techniques like identity normalization.