Abstract: In order to solve the problems that exist in 3D segmentation reconstruction given a series of images from calibrated cameras, a variational method based on probabilistic formulation was proposed. First, through computing most probable surface that gave rise to the images, a 3D surface consistent with these segmentations was built. Then, through fusing joint probabilities, the mean intensity and variance of the extracted object and background were reconstructed. At last, by using a level set framework, the numerical implementation of surface energy equation was carried out. The proposed method can reconstruct complex topologies and cope with noisy data. Compared to carving techniques and stereoscopic segmentation, the experimental results show the effectiveness and robustness of the method with segmentation and reconstruction of arbitrary 3D objects.

Key words: Multiple-views scenes, 3D objects, Bayesian inference, Variational segmentation, Level set

摘要： 针对校准摄像机采集系列图像的三维分割重构问题，提出了一种新的面向概率描述的变分方法。首先，计算系列图像的极大似然曲面，可重构与分割保持一致的三维曲面；接着，融合联合概率，可重构目标对象及图像背景的平均强度及标准差；最后，采用水平集框架，可实现对曲面能量方程的数值模拟。该方法适用于复杂拓扑结构重构及噪声数据处理。实验结果表明，该方法实用性好，鲁棒性强，对任意三维对象的分割重构效果较形状雕刻方法及体视分割方法理想。

关键词: 多视场景, 三维目标, 贝叶斯推理, 变分分割, 水平集