Group-wise relation mining for weakly-supervised fine-grained multimodal emotion recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Advanced multimedia technologies enable fine-grained emotion recognition, but traditional methods face challenges with costly annotations and performance gaps in self-supervised learning.
• Existing weakly-supervised methods struggle with fine-grained emotion recognition due to ambiguous segment boundaries from coarse annotations.

Purpose of the Study:

• To propose a group-wise relation mining method (GRM4WFER) for enhanced fine-grained emotion recognition under weak supervision.
• To improve the discriminability of segment boundaries in challenging samples by broadening the information receptive field.

Main Methods:

• GRM4WFER jointly explores relationships between segments across multiple trials of the same subject.
• Integrates a relative distance constraint with Bayesian probabilistic encoding to enhance boundary discriminability.
• Employs relation reasoning strategies for adaptive exploration of semantic contexts with an extended receptive field.

Main Results:

• The proposed GRM4WFER model demonstrates superior performance compared to state-of-the-art baselines.
• Experiments conducted on two fine-grained multimodal emotion datasets validate the model's effectiveness.

Conclusions:

• GRM4WFER effectively addresses the limitations of traditional and existing weakly-supervised methods for fine-grained emotion recognition.
• The method offers a promising approach for precise emotion recognition in complex, real-world scenarios.