Driver Head Analysis Based on Deeply Supervised Transfer Metric Learning with Virtual Data

摘要

Driver head analysis is of paramount interest for the advanced driver assistance systems (ADAS). Recently proposed methods almost rely on training with labeled samples, especially deep learning. However, the labeling process is a subjective and tiresome manual task. Even trickier, our application scene is driver assistance systems, where the training dataset is more difficult to capture. In this paper, we present a rendering pipeline to synthesize virtual-world driver head pose and facial landmark dataset with annotation by computer 3D animation software, in which we consider driver's gender, dress, hairstyle, hats and glasses. This large amounts of virtual-world labeled dataset and a small amount of real-world labeled dataset are trained together firstly by deeply supervised transfer metric learning method. We treat it as a cross-domain task, the labeled virtual data is a source domain and the unlabeled real-world data is a target domain. By exploiting the feature self-learning characteristic of deep networks, we find the common feature subspace between them, and transfer discriminative knowledge from the labeled source domain to the labeled target domain. Finally we employ a small number of real-world dataset to fine-tune the model itera-tively. Our experiments show high accuracy on real-world driver head images.