Gaze Object Detection Based on 3D Body Posture

Gaze object detection in computer vision is challenging, particularly in scenes with motion blur, multiple overlapping objects, and unclear object boundaries. Traditional methods that primarily rely on head movements struggle to perform well in such complex environments. To address these challenges, our dual-attention multimodal architecture enhances accuracy by establishing a connection between the body orientation and the gaze object while also incorporating depth information for a more precise analysis. Additionally, we introduce a novel IoC (Intersection over Confidence) metric, which provides better differentiation between gaze objects and non-targets. Our approach demonstrates superior performance on both the GOO and Gazefollow datasets, outperforming existing methods in these challenging conditions.