Estimating gaze direction is a key task in computer vision, especially for understanding where a person is focusing their attention. It is essential for applications in assistive technology, medical diagnostics, virtual environments, and human–computer interaction. In this work, we introduce Dual Focus-3D, a novel hybrid deep learning architecture that combines appearance-based features from eye images with 3D head orientation data. This fusion enhances the model’s prediction accuracy and robustness, particularly in challenging natural environments. To support training and evaluation, we present EyeLis, a new dataset containing 5206 annotated samples with corresponding 3D gaze and head pose information. Our model achieves state-of-the-art performance, with a MAE of 1.64° on EyeLis, demonstrating its ability to generalize effectively across both synthetic and real datasets. Key innovations include a multimodal feature fusion strategy, an angular loss function optimized for 3D gaze prediction, and regularization techniques to mitigate overfitting. Our results show that including 3D spatial information directly in the learning process significantly improves accuracy.

Gaze interpretation is a fundamental aspect of social communication, especially for children with autism spectrum disorder (ASD), who frequently encounter many difficulties in social interactions. Despite the considerable advances made in gaze tracking technologies, such as those based on RGB and RGB-D, the accurate measurement of gaze direction remains a significant scientific challenge. This paper proposes a novel approach utilizing the Azure Kinect to improve the measurement of gaze intention in children with ASD, providing accurate estimations of both gaze direction and head position. To evaluate the effectiveness of the proposed methodology, an experimental trial was conducted with eight participants of varying statures. The intersection of the estimated gaze with the target plane was also analyzed throughout 38-min sessions. The results demonstrated high accuracy, with a minimum angular error of 2.5° using pupil positions, 2.06° using head orientation, and average errors of 4.46° and 3.19°, respectively. This approach was tested on a dataset of children with ASD to track their gaze towards a clinician, as this information is essential for assessing the children’s social intent and interactions with others, facilitating more precise clinical assessments for children with autism.