OGaze: Gaze Prediction in Egocentric Videos for Attentional Object Selection

This paper proposes a novel gaze-estimation model for attentional object selection tasks. The key features of our model are two-fold: (i) usage of the deformable convolutional layers to better incorporate spatial dependencies of different shapes of objects and background, (ii) formulation of the gaze-estimation problem in two different ways, i.e. as a classification as well as a regression problem. We combine the two different formulations using a joint loss that incorporates both the cross-entropy as well as the mean-squared error in order to train our model. The experimental results on two publicly available datasets indicates that our model not only achieved real-time performance (13–18 FPS), but also outperformed the state-of-the-art models on the OSdataset along with comparable performance on GTEA-plus dataset.