DualGazeNet: robust gaze estimation via dual-stream disentanglement and hybrid supervisions

Intelligent eye-tracking technology plays a crucial role in visual-function training, vision rehabilitation, and cognitive/motor-control training. Due to the eye’s distinctive appearance as well as the considerable variability in head pose and gaze direction, existing appearance-based gaze estimation approaches still face performance limitations. To improve robustness in unconstrained settings, this paper proposes DualGazeNet. DualGazeNet employs a ResNet50 backbone coupled with a dual-branch fully-connected architecture that explicitly disentangles yaw and pitch learning. This paper innovatively designs a hybrid multi-loss function that jointly supervises angle classification and regression tasks, and introduces multi-scale feature fusion together with a spatialattention mechanism to strengthen feature representation. Evaluated on the MPIIGaze and Gaze360 datasets, DualGazeNet achieves state-of-the-art performance, validating the effectiveness of our approach.