Object Modeling and Recognition from Sparse, Noisy Data via Voxel Depth Carving

In this work, we make the case for using volumetric information for shape reconstruction and recognition from noisy depth images for robotic manipulation. We provide an efficient algorithm, Voxel Depth Carving (a variant of Occupancy Grid Mapping) which accomplishes this goal. Real-world experiments with lasers, RGB-D cameras, and simulated sensors in both 2D and 3D verify the effectiveness of our algorithm in comparison to traditional point-cloud based methods.