WEyeDS: A desktop webcam dataset for gaze estimation

Anatolii Evdokimov, Catherine Finegan-Dollak, Arryn Robbins

Abstract: We introduce the Webcam Eye Dataset (WEyeDS) 1, a dataset of images collected from desktops and laptops. WEyeDS will be openly available with the goal of providing researchers a dataset for training gaze estimation models. Our preliminary dataset was collected from 38 participants who completed prosaccade tasks on their personal desktop or laptop computer. We benchmarked the dataset using a convolutional neural network. The results of our benchmark model demonstrate some challenges in accuracy due to variability across devices and small sample size. Future work will significantly expand the number of participants and frames.