# EmotiEffNet Facial Features in Uni-task Emotion Recognition in Video at ABAW-5 competition

**Authors:** Andrey V. Savchenko  
**arXiv:** [2303.09162](https://arxiv.org/abs/2303.09162)  
**Date:** 2023-03-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.09162v1/

## Key points

- Study focuses on using pre-trained convolutional networks from the EmotiEffNet family for frame-level feature extraction
- Authors propose an ensemble approach combining a multi-layered perceptron with a LightAutoML-based classifier
- Ensemble model used for facial expression recognition and action unit detection
- Post-processing techniques implemented to improve accuracy for sequential frames
- Experimental tests conducted using the Aff-Wild2 database
- Proposed model achieves significantly higher macro-averaged F1-scores compared to baseline models for facial expression recognition and action unit detection
- Improved concordance correlation coefficients for valence/arousal estimation
- EmotiEffNet facial features and ensemble approach show effectiveness in emotion recognition in video analysis
- Methodology can contribute to advancements in understanding affective behavior analysis in real-world scenarios and provide more accurate results than existing approaches.

## Summary

In this article titled "EmotiEffNet Facial Features in Uni-task Emotion Recognition in Video at ABAW-5 competition," Andrey V. Savchenko presents the results of their team's participation in the fifth Affective Behavior Analysis in-the-wild (ABAW) competition. The study focuses on exploring the use of pre-trained convolutional networks from the EmotiEffNet family for frame-level feature extraction. The authors propose an ensemble approach that combines a multi-layered perceptron with a LightAutoML-based classifier. This ensemble model is used for facial expression recognition and action unit detection. Additionally, post-processing techniques are implemented to smooth the results and improve accuracy for sequential frames. To evaluate the performance of their model, experimental tests are conducted using the large-scale Aff-Wild2 database. The results demonstrate that their proposed model achieves significantly higher macro-averaged F1-scores for facial expression recognition and action unit detection compared to baseline models. Furthermore, it also shows improved concordance correlation coefficients for valence/arousal estimation. Overall, this study highlights the effectiveness of using EmotiEffNet facial features and an ensemble approach for emotion recognition in video analysis. The findings suggest that this methodology can contribute to advancements in understanding affective behavior analysis in real world scenarios and provide more accurate results than existing approaches.