Combining Local and Global Features for 3D Face Tracking

Abstract

In this paper, we propose to combine local and global features in a carefully designed convolutional neural network for 3D face alignment. We firstly adopt a part heatmap regression network to predict the landmark points on a local granularity by generating a series of heatmaps for each 3D landmark point. To enhance the ability of local feature representation, we incorporate the designed network with a part attention module, which transfers the convolution opereation into a channelwise attention opereation. Additionally, we take all these heatmaps alongside the input image as the input of another shape regression network in order to model the feature representations from local discrete regions to a global semantically continuous space. Extensive experiments on challenging datasets, AFLW2000-3D, 300VW and the Menpo Benchmark, show the effectiveness of both the global consistency and local description in our model, and the proposed algorithm outperforms state-of-the-art baselines.