Precise eye localization with improved SDM

A joint cascaded framework for simultaneous eye detection and eye state estimation

“…Some other experiments are conducted given the face or specific eye region. In this paper, another comparison experiment with a similar work [13] is conducted. [13] improves the basic SDM[14] by extracting SIFT features at first stage and 330 LBP at the following iterations.…”

“…Yang et al [12] propose to detect the pupil by using different Gabor kernels to convolute with the image, which highlights the eye-and-brow regions. By employing a coarse to fine strategy for robust initial-75 ization, Zhou [13] propose multi-scale nonlinearet al feature mapping based on the Supervised Decent Method (SDM) [14] for eye detection. They use 14 eye related key points to capture the contextual information.…”

Hybrid eye center localization using cascaded regression and hand-crafted model fitting

“…Subsequently, they use a cascade of regression forests with binary pixel difference features to estimate the eye centers. Inspired by the recent success of the SDM method for facial feature alignment Zhou et al [28] propose a similar method for eye center localization. Unlike the original SDM work, their regressor is based on a combination of SIFT and LBP features.…”

“…We find that using HoG is especially helpful for bright eyes, where variation in appearance due to different lighting and image noise is more apparent, hurting the performance of regressors employing simple pixel difference features. Zhou et al [28], also employ advanced image features, but in contrast to them we use regression forests at each level of our cascade. Finally, while [17,21] estimate eye center positions independently, we find that due to the large amount of correlation between the two eyes it is beneficial to estimate both eyes jointly.…”

“…Our first contribution is a new method for eye center localization that employs complex features and complex regressors. It outperforms simple regressors with complex features [28] and complex regressors with simple features [17,21]. Similar to [17,21] our method is based on cascaded regression trees, but unlike these authors, following [6,18,14,26], our features for each cascade are anchored to the current eye center estimates.…”