Fast model-based 3D human face reconstruction from silhouettes of multiple views

Abstract

This thesis presents the studies of developing a new procedure for reconstructing 3D human face model from multiple silhouettes images derived from photographs that are taken from different views. Many techniques have been reported in the literatures for reconstructing 3D face model from photographs. Some techniques require accurate feature detectors to extract distinguishable facial features to perform reconstruction. Others use the global appearance of face as the cue for the shape. However, the performance of these methods is subject to the illumination settings of the image acquisition environment. It is therefore essential to develop a technique that is less sensitive to the environment setting. Camera calibration is critical to the accuracy of the reconstruction result and a planar chessboard-like pattern has been used for calibration in order to reduce the error introduced by the initial alignment. An analytical solution for solving both the intrinsic and extrinsic parameters of the camera position has been reviewed. In order to reduce the sensitivity to the change of illumination setup, segmentation with spatial constrain has been applied to image represented in CIELAB color space to effectively separate the face from the background. A high resolution 3D Principal Component Analysis (PCA) model derived statistically from 100 individuals is used to model human face. By varying the PCA model parameters, the 3D face model can be reconstructed so that its projections are consistent with the input silhouettes extracted from the images. Since a large part of the head image is covered by hair and it does not convey any information related to the face, the face model used should not be a full human head model. The new model fitting technique is different from previous works in that it only requires the knowledge of face boundary that is covered by the face model rather than the entire head image area. A vertex-driven approach has been developed for searching the parameter vector of the PCA face model. As the approach only depends on the vertices that are located on the boundary of the model projection, an implicit but significant down-sampling is introduced to help reducing the computational burden. The parameter vector is updated in an iterative manner according to a distance metric between the model projection and the input silhouette. With the high resolution model constituting of more than 70,000 vertices, the computational efficiency of new approach has been well demonstrated and the reconstruction process requires less than a minute to converge. Different combinations of the control parameters have been examined to achieve the best performance. Experimental results have demonstrated that the accuracy of the proposed reconstruction procedure is high and with fast convergent rate.