Sparse-Region Net: Local-Enhanced Facial Depthmap Reconstruction from a Single Face Image

摘要

In this paper, we propose a novel end-to-end deep neural network for region-enhanced depthmap reconstruction from a single face image. Unlike most of popular depthmap reconstruction methods, the proposed network fully takes the region information of RGB images into consideration, and thus results in more accurate correspondences. Specially, we treat the depthmap reconstruction task as an independent problem, investigating to improve facial depth regression through multitask learning (facial landmark detection and mask detection), and then a attention region training technique to utilize the rich texture information in some semantic regions to infer local detailed depth information. Next, we describe a zigzag dilated convolution framework to sparse the kernel size of our network, which will enlarge the network's receptive field and simultaneously avoid gridding artifacts. In training, we also propose a new method for setting the learning rates, called discrete cyclical learning rates, to improve the results. Experimental results illustrate that the proposed Sparse-Region Net (SRN) outperforms the state-of-the-art baseline methods by a large margin.