Abstract:

Many applications in computer vision and human-computer interaction need the determination of the users' area of interest. The groundbreaking deep learning discoveries have received a lot of attention in the gaze estimate literature. The transition of gaze estimation systems from single-user confined settings to multi-user unconstrained environments has been made possible by deep learning algorithms' ability to be deployed in complicated unconstrained situations with considerable volatility. In a number of disciplines, including security, psychology, computer vision, and medical diagnostics, eye tracking is swiftly emerging as a very important tool. Security apps also employ gaze to evaluate suspicious gaze behavior. In educational institutions, automated eye gaze analysis of students during tests is a use case that might lessen cheating. The main focus of this paper is the research and investigation of several CNN architectures for gaze estimation and prediction. In this study, two tasks—gaze estimation and gaze prediction based on known gaze-points—have been devised. Several CNNs were used in the first challenge to find the most accurate gaze estimation. Using the previously calculated gaze vectors and the spatiotemporal information contained in previously collected eye-image sequences, we anticipate gaze positions in the second challenge. To predict the locations of the next view, we used a Long Short Term Memory (LSTM), Transformers based on self-attention, and positional encoding.

Description:

AI, computer vision, machine learning, simulation, eye tracking, human recognition , eye gaze estimation