Binocular Stereo Vision Based Depth Estimation And Applications

Binocular stereo vision based depth estimation is a fundamental problem in computer vision.It has important applications in 3D face reconstruction and recognition,human pose estimation,and 3D scene reconstruction.In this thesis,the binocular depth estimation task is studied intensively.In addition,the binocular depth estimation is applied to face spoofing detection task and corresponding application is also studied.Aiming at producing more accurate disparity predictions in thin structures or near boundaries,an edge detection integrated stereo matching network EdgeStereo is designed,which achieves state-of-the-art performance on several stereo matching benchmarks.The model consists of an edge detection sub-network and a stereo matching sub-network with two efficient and effective modules embedded: a context pyramid to encode multi-scale context cues for disparity estimation;a residual pyramid formulating disparity prediction and refinement conducted in a single decoder structure.In order to conduct multitask learning,an edge-aware disparity smoothness loss is proposed,furthermore edge detection task and stereo matching task are demonstrated helpful for each other based on the proposed EdgeStereo model.In this paper,the combination of binocular depth feature and appearance feature based face spoofing detection is also studied,and correspondingly two methods are proposed.The first method is the discriminative representation combination of template face matched binocular depth feature and spatial pyramid coding micro-texture feature,the difference in stereo structure for real and fake faces are captured by designed template face registration algorithm,and the complementarity of binocular depth feature and appearance feature is utilized to make liveness judgment;in order to further improve the robustness for face spoofing detection,the second method called the discriminative representation combination of deep learning based detection network and spatial pyramid coding micro-texture feature is proposed,achieving state-of-the-art performance on face spoofing detection databases.