A Novel Rotation-aware Representation Learning Procedure for Gaze Estimation to Assist Disabled Individuals in Communicating with the World Through Various Applications

Overview

The human eyes provide an essential way of taking in visual information from the world. Fascinatingly, they can also be used to indicate a wide range of messages or emotions. Since moving the eyes requires minimal effort, even quadriplegic individuals can communicate with their eyes. Gaze estimation, the task of inferring the direction where an individual’s eyes are pointed to, provides these individuals with a refined method of connecting with the world. However, gaze estimation tasks lack comfortable real-world applications because they are inherently challenging and often unreliable due to the wide variations in the characteristics of each pair of eyes. In my research, I propose a representation-learning-based training procedure incorporating a rotation matrix to increase the accuracy of gaze estimation tasks. Consequently, the proposed approach exponentiates the disentanglement of gaze-related latent features from the rest, ultimately achieving higher accuracy in more robust, harder samples. The Few Shot test results highlight an average increase of 16.5% in accuracy compared to previous state-of-the-art methods on GazeCapture dataset. Ultimately, my findings will provide disabled individuals with vast applications to more comfortably communicate with the world. For instance, my prototype of a gaze-controlled wheelchair allows paralyzed individuals to become physically independent once again.