Multi-task Learning for Structured Output Prediction

Facial landmark detection is an important step for many perception tasks. In this paper, we address facial landmark detection as a structured output regression problem, where we exploit the strong dependencies that lie between the facial landmarks. For this, we propose a generic multi-task regression framework for structured output problems. The learning of the output structure is achieved through a regularization of the supervised task, in an unsupervised way. Therefore, the proposed framework allows the use of unlabeled input and/or label only output data. In this article, the formulation is instantiated as a deep architecture, and evaluated on two public challenging datasets: LFPW and HELEN. We show that our regularization scheme improves the generalization of deep neural networks, and accelerates their training. The use of unlabeled data is also explored, showing an additional improvement of the results. An opensource implementation of our approach is provided. ∗Corresponding author Email address: [email protected] (Soufiane Belharbi) Preprint submitted to Pattern Recognition November 21, 2016 ar X iv :1 50 4. 07 55 0v 4 [ cs .L G ] 1 8 N ov 2 01 6