A Crux on Audio-Visual Emotion Recognition in the Wild with Fusion Methods

초록

This paper introduces recent advancements in emotion recognition using multimodal approaches, with a particular focus on real-world data. The primary challenge in this domain is accurately analyzing emotions in uncontrolled environments, where factors such as occlusion, illumination, and noise complicate data interpretation. To address these issues, we review studies that utilize various fusion techniques-such as early fusion, intermediate fusion, and late fusion-to effectively combine multiple modalities. Additionally, we conducted experiments using real-world dataset to evaluate performance and uncovered key limitations, including contextual mismatches and challenges in integrating diverse modalities. These findings highlight the importance of developing robust fusion strategies for real-world emotion recognition systems. Future work should focus on optimizing fusion strategies and enhancing model robustness to better handle the complexities of real-world data, and advance the field of Human-Computer Interaction. © 2025 IEEE.

키워드