单幅图像真实感虚拟试戴技术

Real-Like Virtual Fitting for Single Image

提出了一种针对眼镜、帽子等头部饰品的单幅图像真实感虚拟试戴技术,其关键在于虚拟饰品的三维注册和虚实图像的合成.首先提出了一种将人脸关键点检测与刚体姿态估计相结合,求解单幅图像中人脸三维注册信息的算法.然后阐述了借助像素颜色混合和深度缓冲检测技术解决虚实图像合成中遮挡关系和模型材质问题的方法.在AFLW(Annotated Facial Landmarks in the Wild)人脸数据库上对三维注册算法进行了量化测评,结果表明该算法的精度满足虚拟试戴技术的要求.在较大角度姿态变化以及部分遮挡条件下的实验结果表明,提出的虚拟试戴技术快速准确,试戴效果自然逼真.

An image based real-like virtual fitting system focusing on head accessories such as glasses and hat is proposed in this paper. The key techniques are 3D registration and virtual-real synthesis. Firstly, a facial landmark detection and pose estimation based algorithm for capturing registration data is presented. Then the approach using color blending and depth buffering to solve occlusion and model material problem is discussed. The registration algorithm is evaluated on the AFLW（Armotated Facial Landmarks in the Wild） dataset and testified to be precise enough for virtual fitting. Finally, a series of virtual fitting results are showed. The experimental results, which are obtained on the condition of pose variation and part occlusion, indicate that the proposed virtual fitting technique is fast, accurate and real-like.