The paper addresses the challenge of introducing unlabeled facial images into supervised AU detection frameworks. It introduces a novel approach that jointly learns AU domain separation, reconstruction, and facial landmark detection. By sharing parameters between tasks, the proposed framework demonstrates superior performance compared to state-of-the-art methods on two benchmarks.
The introduction of multi-task learning strategies aims to address labeling challenges and additional domain shifts due to pose variations and occlusions. The feature alignment scheme based on contrastive learning enhances the reconstruction process by adding intermediate supervisors for improved feature alignment. Extensive experiments validate the effectiveness of the proposed method for AU detection in diverse scenarios.
Experimental results showcase significant improvements over traditional methods in both source and target domains. The framework's strong generalization performance makes it applicable for various applications like human-computer interaction, emotion analysis, and car-driving monitoring.
A otro idioma
del contenido fuente
arxiv.org
Ideas clave extraídas de
by Ziqiao Shang... a las arxiv.org 03-05-2024
https://arxiv.org/pdf/2310.05207.pdfConsultas más profundas