A Real Sense 3D Face Reconstruction System Based on Multi-view Stereo Vision

摘要

This paper proposed a system for a real 3D facial reconstruction method based on multi-view stereo vision generated using an orthographic 3D model. Multi-view stereopsis is an effective technology for expanding perspective and reducing noise. The presented multi-view stereo vision is implemented as a multi-view stereo vision using calibration, stereo matching, combining, reconstruction and texture mapping procedures. This paper makes a systemic contribution and two technical contributions. Its systemic contribution is that it demonstrates a state-of-the-art passive stereo vision system for meaningful orthographic 3D facial reconstruction. We have tested our method on various subjects, including two actors and an artificial plastic human head model. The primary technical contribution is an algorithm that combines stereo vision data based on clustering, which includes the following procedures: outlier detection, filling and data combining. The second contribution is a multi-view calibration method, which results in an orthographic position model. An orthographic position is an important requirement and is useful in various medical fields, e.g., maxillofacial surgery research. In addition, we describe a cluster sampling algorithm for undersampling point clouds. The presented sampling algorithm keeps data on an average distribution. A quantitative evaluation proves that the system is an effective, low-cost and high-resolution solution to reconstructing a 3D face model without markings or structured light.