Gaze Estimation Based on Difference Residual Network
Bei Yan, Xiangyu Tang · 2021 International Conference on Computer Information Science and Artificial Intelligence (CISAI) · 2021
Because of the variabilities in eye shapes and inner eye structures among individuals, the universal models that regress gaze directions directly from a single face or eye image through the neural network obtain limited accuracies. There is a person-specific gaze bias caused by the kappa angle between the model’s predicted and the actual gaze direction. Currently, people-specific models are usually established through calibration to increase accuracy. This paper proposes a novel estimation method of gaze direction based on the Differential Residual Network (D-ResNet),Which works by training the D-ResNet to predict gaze difference between two eye images (the same eye of the same subject) to eliminate gaze bias. At test time, only a few person - specific calibration samples are needed to infer the new eye image’s gaze direction. Our model also incorporates the head pose vector to improve the robustness to free-head. Moreover, our left and right eye models are trained and tested separately which can select the dominant eye model based on the test error to predict gaze. Experiments on 2 public datasets validate our approach outperforms state-of-the-art person-independent and person- specific gaze estimation models. In particular, our model doesn’t fail to calibrate even if there is only one calibration sample.