Direction-Sensitivity Features Ensemble Network for Rotation-Invariant Face Detection

Abstract

Recent deep learning based Rotation-Invariant Face Detection (RIFD) algorithms make efforts to explore a mapping function from face appearance to the rotation-in-plane (RIP) orientation. Most methods propose to predict RIP angles in a coarse-to-fine cascade regression style and improve the overall RIFD performance. However, the problem of suboptimal between the models of training phase and testing phase cannot be solved because of its cascaded nature. The weakness of ambiguous mapping between face appearance and its real orientation would also degrade the performance considerably. In this paper, we propose a novel Direction-Sensitivity Features Ensemble Network for rotation-invariant face detection (DFE-Net) which learns an end-to-end convolutional model for RIFD from coarse to fine. Specifically, the incline bounding box regression is implemented by introducing angle prediction based on improved SSD. A Direction-Sensitivity Features Ensemble Module (DFEM) is adopted in the network to progressively focus on the awareness of face angle information, which can learn and accurately extract features of rotated regions and locate rotated faces precisely. Finally, we add multi-task loss to guide the learning process to captures consistent face appearance-orientation relationships. Extensive experiments on two challenging benchmarks demonstrate that the proposed framework achieves favorable performance and consistently outperforms the state-of-the-art algorithms.