抄録
Various gaze estimation methods have been proposed, but they have not been evaluated well in random
environment, and their estimation accuracy are not sufficient. We developed a gaze estimation method using
convolutional neural network based on left and right eye images. In addition, we evaluated the present method with a cross-dataset evaluation using MPIIGaze and performed cross-validation using a dataset of eye images obtained in unrestricted daily settings. The results of this study are summarized as followings. First, we created a learning machine model that estimates the gaze direction of each of the left and right eyes from both eyes images, using a convolutional neural network. The accuracy was 10.3°, which was equal to or greater than the previous excellent gaze estimation method. Second, we compared the model with the input of both eyes and the model with the input of one eye. As a result, it was found that the accuracy of the model using both eyes images was 54.2% better than the accuracy of the model using one eye image. Furthermore, it was found that the accuracy especially in the horizontal direction was improved. Finally, in the cross-dataset evaluation, the accuracy in the vertical direction is better than the horizontal direction, which is different from the result that the horizontal direction is generally better than the vertical direction. On the other hand, the within-dataset evaluation showed the opposite result. It is indicated that the vertical estimation in the proposed method is more flexible to environmental variations than the horizontal estimation.