Automated Tool for Infant Facial Action Unit Detection

University of Pittsburgh researchers have developed an automated tool, named Infant AFAR, for detecting facial actions units (AUs) in infant faces, addressing the limitations of manual BabyFACS annotation and the need for an automated, real-time system.

Description

Infant AFAR utilizes a convolutional neural network (CNN) approach, pre-trained on adult faces and fine-tuned with well-annotated infant databases, to tackle the unique characteristics of infant faces. The researchers evaluated Infant AFAR's performance within databases, age groups, and cross-database and cross-age-group scenarios, finding that it generally outperforms or performs comparably to within-database performance. The research also makes Infant AFAR publicly available as an extension of AFAR, enabling validated facial action unit detection in infants and young children.

Applications

· Social communication research
· Infant development research
· Developmental disorders research

Advantages

AU detectors trained with adult faces perform poorly in detecting AUs in infant faces, highlighting the need for models specific to infant faces. The proposed Infant AFAR achieves state-of-the-art performance for infant AU detection. For adults, AU occurrence and intensity detection are enabled for 12 action units (AU). The AU chosen were selected on the criterion that they have base rates greater than 5% in BP4D+. Action unit intensity estimation is enabled for 5 of these AU. For infants, AU occurrence is enabled for 9 action units that are involved in expression of positive and negative affect.

Invention Readiness

This invention is in the software stage of development.

IP Status

Software