Eye landmarks detection via two-level cascaded CNNs with multi-task learning

Despite extensive researches for eye localization, accuracy of eye landmarks detection is easily influenced by illumination and the difference in eye state conversion between open and close. To address these problems, we present a novel approach for eye landmarks detection with two-level cascaded convolutional neural networks. Network at the first level utilizes eye state estimation as auxiliary task to provide initial positions of the eye. The shallower network at the second level fine tunes eye positions by taking small regions centered at predicted eye points as input. To train our model, we introduce OCE dataset, the first dataset with the eye in different states. Our method achieves mean detection error of 5.6% on OCE dataset. Further experiments are tested on UBIRIS V1.0, MMU V1.0 and MICHE-I and their results demonstrate acceptable performance of our method.