Voxel coloring needs accurate silhouette information of foreground objects for the reliable 3-D scene reconstruction of the objects.
This thesis proposes a new method to optimally detect the foreground objects through a statistical model of the color difference between two scenes with and without the foreground objects, respectively. This thesis shows that image noise always does not satisfy the conventional zero-mean Gaussian assumption. Therefore, we propose a new noise model of each color channel for precisely estimating the statistics of background Euclidean color distances. Consequently, we can find the optimal threshold to classify the foreground objects from the background.
Even though we use the optimal threshold, there can be inevitable false classifications. To reject these erroneous cases, we adopt the Graph Cuts that efficiently minimizes the global energy. However, there still exists the opportunity of false classification when the brightness change in the shadow region is increased. To resolve this problem we use the property of the shadow region that the chromaticity remains constant while the brightness changes in shadow region.
We analyze our proposed methodology through various experiments and have shown the feasibility of reliable extraction of silhouettes and shadow. Consequently, we can reconstruct the 3D shape more accurately than by the conventional thresholding-based method.
