# Aff-Wild2: Extending the Aff-Wild Database for Affect Recognition

**Authors:** Dimitrios Kollias, Stefanos Zafeiriou  
**arXiv:** [1811.07770](https://arxiv.org/abs/1811.07770)  
**Date:** 2018-11-11  
**Summary page:** https://www.summarizepaper.com/en/paper/1811.07770v1/

## Key points

- The paper discusses the problem of automatic understanding of human affect using visual signals
- The authors propose using the circumplex model of affect to capture and describe emotions in terms of valence and arousal
- The Aff-Wild database was developed as a large-scale "in-the-wild" database with around 1,200,000 frames
- The authors extend the Aff-Wild database by adding 260 more subjects and 1,413,000 new video frames, creating Aff-Wild2
- Aff-Wild2 videos exhibit variations in pose, age, illumination conditions, ethnicity, and profession
- Deep neural architectures based on joint training of convolutional and recurrent neural networks with an attention mechanism are developed
- Database-specific and cross-database experiments are performed using Aff-Wild2 along with the RECOLA database
- Results show promise for utilizing Aff-Wild2 dataset and deep neural architectures for analyzing human behavior in terms of continuous emotion dimensions

## Summary

The paper titled "Aff-Wild2: Extending the Aff-Wild Database for Affect Recognition" by Dimitrios Kollias and Stefanos Zafeiriou discusses the problem of automatic understanding of human affect using visual signals. Over the past 20 years, there has been significant interest in this area; however, human emotional states are complex and require expressive emotional descriptors to capture and describe their complexity. To address this issue, the authors propose using the circumplex model of affect, which describes emotions in terms of valence (positive or negative) and arousal (power of emotion activation). Recent progress in emotion recognition has been made through deep neural architectures and large training databases. The Aff-Wild database was developed as the first large-scale "in-the-wild" database with around 1,200,000 frames. In this paper, the authors extend the Aff-Wild database by adding 260 more subjects and 1,413,000 new video frames. This extended dataset is called Aff-Wild2. The videos in Aff-Wild2 are downloaded from YouTube and exhibit variations in pose, age, illumination conditions, ethnicity, and profession. The authors perform both database-specific and cross-database experiments using Aff-Wild2 along with the RECOLA database. They develop deep neural architectures based on joint training of state-of-the-art convolutional and recurrent neural networks with an attention mechanism. This approach allows them to exploit both invariant properties of convolutional features and model temporal dynamics present in human behavior through recurrent layers. The results obtained from these experiments show promise for utilizing the extended Aff-Wild dataset as well as the developed deep neural architectures for visual analysis of human behavior in terms of continuous emotion dimensions. Overall, this paper contributes to advancing research in affect recognition by providing a larger dataset (Aff-Wild2) that captures real-world variations in human emotional states. The proposed deep neural architectures demonstrate improved performance in analyzing human behavior based on continuous emotion dimensions.