Facial landmark detection transfer learning for a specific user in driver status monitoring systems

The wide variety of human faces make it nearly impossible to prepare a complete training data set for facial landmark detection. Because of this, the performance of facial landmark detection is unlikely to be sufficient for driver status monitoring (DSM) systems. To improve the performance for a specific person (SP) by collecting data about that person, we propose the generator and discriminator model using the Lucas-Kanade assistance (GDA) algorithm for compiling a training data set. Even when data for a specific user can be collected, another issue is how to efficiently, effectively, and quickly re-train the model using an insufficient data set. To address this problem, we propose a novel method of transfer learning in the context of composite backbone networks (GBNet). The assistant backbone of GBNet is trained on a large unspecified people (USP) data set in the source domain and transfers its representation to the lead backbone, which is trained by a small SP data set in the target domain. In addition, we design an assistance loss function with output that is not only close to the SP data set, but also consistent with a USP data set with respect to labeled images. We test the proposed method using the 300 Videos in the Wild (300VW) data set and our own data set. Furthermore, show that the proposed method improves the stability of predictions. We expect our method to contribute to the realization of stable DSM systems.