Deep Cascaded Regression for Face Alignment

We propose a novel cascaded regression framework for face alignment based on a deep convolutional neural network (CNN). In most existing cascaded regression methods, the shape-indexed features are either obtained by hand-crafted visual descriptors or by leaning from the shallow models. This setting may be suboptimal for the face alignment task. To solve this problem, we propose an end-to-end CNN architecture to learn highly discriminative shape-indexed features. First, our deep architecture encodes the image into high-level feature maps in the same size of the image via three main operations: convolution, pooling and deconvolution. Then, we propose "Shape-Indexed Pooling" to extract the deep features from these high level descriptors. We refine the shape via sequential regressions by using the deep shape-indexed features, which demonstrates outstanding performance. We also propose to learn the probability mask for each landmark that can be used to choose the initialization from the shape space. Extensive evaluations conducted on several benchmark datasets demonstrate that the proposed deep framework shows significant improvement over the state-of-the-art methods.