A multi-task collaborative learning method based on auxiliary training and geometric constraints

Abstract

In facial landmark detection, cascaded deep convolutional neural networks have high model complexity, inadequate training of underlying parameters and difficult network initialization. To address these problems, a multi-task collaborative learning method is proposed based on auxiliary training and geometric constraints. Firstly, by virtue of the correlation between facial landmark localization and head pose estimation, a deep network model was designed based on joint optimization of these two tasks to simultaneously estimate the landmark coordinates and pose angles. Secondly, the auxiliary training technique was applied to enhance the feature learning ability of the network model through adding a back-propagation layer. Then, the geometric constraint algorithm was employed to pre-train the model and its related parameters are used for the initialization of the constructed model, in order to make the model to extract the invariant features of pose variation effectively. Finally, the effectiveness of the proposed method was evaluated on the public database 300W. Experimental results showed that the proposed method was better than the single-task and multi-task learning methods, and yielded that the AUC0.2 value of facial landmark detection is 0.1406 and the error rates of head pose estimation in three-dimensional space are 4.75%, 8.39% and 5.75% respectively.