EYE LOCATION AND STATE ESTIMATION BASED ON LANDMARKS

Abstract

Eye location and state estimation are key steps in the preprocessing of biometrics recognition such as iris, sclera and periocular. Eye images captured in the non-cooperative environments often suffer from serious occlusions and complex backgrounds. To solve this problem, this paper proposes a robust and accurate single-stage framework based on eye landmarks to detect eye key points and estimate the left, right and open and closed states of eyes. In order to train and evaluate the proposed model, a new OCE-1000 dataset was created and manually labeled with eight key points, open and close state for left and right eyes of each image. Experimental results show that the proposed model achieves 98% accuracy of landmark location and 97% accuracy of eye state estimation on OCE-1000 dataset.