Personalized face-pose estimation network using incrementally updated face shape parameters

摘要

In this paper, a deep learning method is proposed for human image processing that incorporates a mechanism to update target-specific parameters. The aim is to improve system performance in situations where the target can be continuously observed. Network-based algorithms typically rely on offline training processes that use large datasets, while trained networks typically operate in a one-shot fashion. That is, each input image is processed one by one in the static network. On the other hand, many practical applications can be expected to use continuous observation rather than observation of a single image. The proposed method employs dynamic use of multiple observations to improve system performance. In this paper, the effectiveness of the proposed method adopting an iterative update process is clarified through its implementation in the task of face-pose estimation. The method consists of two separate processes: 1) sequential estimation and updating of face-shape parameters (target-specific parameters) and 2) face-pose estimation for every single image using the updated parameters. Experimental results indicate the effectiveness of the proposed method.