Stable Realization of Virtual Reality by Model Selection

图像的 1.3-D 重建我们提出了一种理论，用于以尽可能高的精度从图像中计算场景和物体的 3-D 结构，并以定量方式评估计算解决方案的可靠性。我们通过将其应用于两幅图像的 3D 重建和光流的 3D 解释，证明了该理论的有效性。我们还展示了如何计算协方差矩阵，该协方差矩阵表征从图像中提取的特征点的准确性，并通过实验证实了它们与其可靠性评估的正相关性。我们进一步分析了绝对位置和比例标准化的影响，以描述计算形状的可靠性。提出了一种新的曲线曲面显示技术，为图像三维重建的可靠性评估提供了理论基础。2.几何模型选择的虚拟现实实现推导了“几何AIC”和“几何MDL”作为从图像中提取的噪声数据进行几何推理的模型选择标准，并给出了信息论解释。我们将这些标准应用于图像拼接，并证明即使捕获的图像之间的重叠很小，也可以稳定地计算图像变换。我们还设计了用于校准虚拟演播室应用的移动相机的最佳参考模式，并通过模拟和真实图像实验证实，可以通过合并几何模型选择来避免相机配置的退化和解决方案的统计波动。此外，我们通过模拟和真实图像实验证明，通过引入几何模型选择，可以显着提高分离独立对象运动和从运动图像中估计对象数量的准确性。

1.3-D reconstruction from imagesWe have presented a theory for computing the 3-D structure of the scene and objects from images with maximum possible accuracy and evaluating the reliability of the computed solution in quantitative terms. We have demonstrated the effectiveness of our theory by applying it to 3-D reconstruction from two images and 3-D interpretation of optical flow.We have also shown how to compute the covariance matrices that characterize the accuracy of feature points extracted from images and experimentally confirmed their positive correlation with their reliability evaluation. We have further analyzed the effects of normalization of the absolute position and the scale for describing the reliability of the computed shape. A new technique for displaying curves and surfaces are also proposed.We have thus given a theoretical foundation to the reliability evaluation of 3-D reconstruction from images.2.Virtual reality realization by geometric model selectionyWe have derived the "geometric AIC" and the "geometric MDL" as model selection criteria for geometric inference of noisy data extracted from images and given an information-theoretic interpretation. We have applied these criteria to image mosaicing and demonstrated that the image transformation can be stably computed even if the overlaps between captured images are small.We have also designed an optimal reference pattern for calibrating a moving camera for virtual studio applications and confirmed by simulations and real-image experiments that the degeneracy of camera configuration and statistical fluctuations of the solution can be avoided by incorporating geometric model selection. Furthermore, we have demonstrated by simulations and real-image experiments that the accuracy of separating independent object motions and estimating the number of objects from their motion images can be dramatically improved by introducing geometric model selection.