Multi-view Reconstruction of Static and Dynamic Scenes

We explore the reconstruction of a three-dimensional scene from multiple images captured from far away viewpoints (wide-baseline camera arrangement). Such an arrangement is required for complex scenes where the visibility flora any one viewpoint is not sufficient to adequately reconstruct the entire scene. Also, such an arrangement reduces the error in triangulation of the features, thereby improving the accuracy of reconstruction. Our emphasis is on algorithms that recover a volumetric model of the scene from calibrated cameras by explicitly modeling and detecting occlusions. We present a brief overview of the state of the art in such methods for multi-view reconstruction. In particular, algorithms based on a probabilistic framework have become quite popular and produce very accurate models. Several such probabilistic volume reconstruction methods will be described. For the dynamic parts of the scene, where an online reconstruction is needed, simpler methods are required. An important case of such scenes is that of walking people in a surveillance scenario. For this case, we present fast online algorithms that recover approximate shape and appearance models and 3D trajectories of people as they walk in the scene. Finally, we address the problem of optimal placement of cameras in order to acquire the best possible image data for the reconstruction of a given scene according to the particular task specification.