3D object reconstruction from a sequence of images using voxel coloring

Image-based reconstruction from randomly scattered views is a challenging problem. We present a new algorithm that extends Seitz and Dyer's Voxel Coloring algorithm for reconstructing a voxelized representation of 3D object from a series of images. Voxel Coloring traverses a discretized 3D space in "depth order" to identify voxels that have a unique coloring, constant across all possible interpretation of the scene. This approach has several advantages over existing stereo and structure-from-motion approaches to scene reconstruction. First, the technique can handle a great magnitude of visibility change. So, the cameras can be positioned far apart without degrading accuracy or run-time. Second, the technique integrates numerous images to yield dense reconstruction without degrading run-time. Unlike Seitz and Dyer's algorithm, ours considers the perspective projection effect on the voxel size. We also present a different search method that traverses the voxels along the projection rays of the images. The new search method optimizes the search process by employing the geometry constraint of the pinhole camera model. In the new search method, the size of the voxels is non uniform depending on the distance from the images. Experimental results for simulated and real image sequences show the efficiency of our algorithm.