Shape distributions on voxel surfaces for 3D object classification from depth images

Abstract

In this work we address the problem of 3D shape based object class recognition directly from point cloud data obtained from RGB-D cameras like the Kinect sensor from Microsoft. A novel shape descriptor is presented, capable of classifying 'never before seen objects' at their first occurrence in a single view in a fast and robust manner. The classification task is stated as a matching problem, finding the most similar 3D model and view from a database of CAD models gathered from the web to a given depth image. We further show how locally sensitive hashing can be easily adapted to implement fast matching against a database of 2500 CAD models with more than 200000 views in 160 categories. This shape descriptor utilizes distributions on voxel surfaces and can be used in various applications: As a pure 3D descriptor for 3D model retrieval, as a 2.5D descriptor for finding 3D models to partial views or as our main indention as a classification system in the home-robotics domain to enable recognition and manipulation of everyday objects. Experimental evaluation against the baseline descriptors on a dataset of real-world objects in table scene contexts and on a 3D database shows significant improvements.