Eye landmarks detection via weakly supervised learning

摘要

Extensive eye researches provide good results when images are captured under constrained environment. However, the accuracy of eye landmarks detection depends on explicit bounding-box of eye regions and drops severely in non-ideal conditions. This paper has proposed a novel weakly supervised eye landmarks detection algorithm with object detection and recurrent learning modules. The former is combined with faster R-CNN and is competent to detect bounding-box of facial components and initial positions of the eye. The recurrent module is employed for eye landmarks refinement using the initial eye shape. The proposed algorithm can augment training data effectively and our specific format data consist of supervised and weakly supervised samples. Supervised samples have ground truth of bounding-boxes, corresponding classification labels and eye landmarks coordinates while weakly supervised data does not have eye landmarks information. Despite trained on facial images, the proposed method can detect eyes in severely occluded or local view of facial images without prerequisites of face alignment. Further experiments are performed on our supervised testing set and some public datasets. Their results demonstrate the robustness and effectiveness of the proposed method. (C) 2019 Elsevier Ltd. All rights reserved.