Initial bounding box estimation methods for volumetric three-dimensional reconstruction

Volumetric models are widely used in various schemes of three-dimensional (3D) reconstruction: silhouette-based reconstruction, multiview stereo, and SLAM (simultaneous localization and mapping). The volume containing the scene to be reconstructed is divided by a regular grid into small cubes called voxels. This volume is usually bounded by an axis-aligned cube (bounding box). The preliminary to 3D reconstruction is to estimate a bounding box. The paper describes four methods of initial bounding box estimation and provides their experimental evaluation and comparison. The presented methods rely on known camera matrices and do not require any prior information about the scene.