[Japanese | Thesis | Researches in Minoh Lab | Minoh Lab]
The volume intersection method reconstructs shapes of target objects from silhouettes with multiple cameras. With stereo vision, textureless objects are different to be handled. In case of laser range finders, laser absorbing materials can not be handled either. Compared with these two methods, the volume intersection method can reconstruct the shapes of objects as long as their silhouettes are extracted. It means that the method copes with more kinds of objects than the other methods do. But there occurs a problem when missing parts of silhouettes, some parts of shapes are missing.
The missing parts of silhouettes occur mainly when objects' color and background's color are similar. We propose a method for silhouette refinement to reconstruct more accurate shapes.
Missing parts of silhouettes should be limited enough to refine them properly. Random pattern background, which has small regions filled with random colors, can help with it.
Missing parts of silhouettes which are limited can be refined by using multiple cameras around a target object. In certain voxels of interest in the 3D object, corresponding pixels are in silhouettes for some cameras, while they are out of silhouettes for the other cameras. Diffuse reflection colors of the voxels are estimated from the cameras that have corresponding pixels in silhouettes. When this color and background's color for the voxels are similar, the cameras tend to miss the corresponding parts of silhouettes. In this case, the corresponding pixels should be included in the silhouettes for those cameras.
Using the random pattern background and the silhouette refinement together, accurate shapes of objects with various colors are obtained.
In our experiment, proper silhouette refinement was confirmed by comparison between a reconstructed shape and a shape from manually extracted silhouettes. The experimental results show the method is capable of refinement silhouette for objects with various colors.