In this paper, 2D cascaded AdaBoost, a novel classifier designing framework, is presented and applied to eye localization. By the term ?2D?, we mean that in our method there are two cascade classifiers in two directions: The first one is a cascade designed by bootstrapping the positive samples, and the second one, as the component classifiers of the first one, is cascaded by bootstrapping the negative samples (please refer to Fig.1). The advantages of the 2D structure include: (1) It greatly facilitates the classifier designing on huge-scale training set; (2) It can easily deal with the significant variations within the positive (or negative) samples; (3) Both the training and testing procedures are more efficient. The proposed structure is applied to eye localization and evaluated on four public face databases, extensive experimental results verified the effectiveness, efficiency, and robustness of the proposed method.