A Novel Approach Utilizing Down Sampled Mesh Clustering and Neural Network Classification

Abstract 3D object recognition is becoming popular since the popularity of 3D models and printing. Often in an extensive database of 3D objects of various types, finding the required object becomes difficult. In the past, various techniques have been proposed to detect and recognize 3D objects from their shape information efficiently. Such methods either use a 3D shape model derived from a 2D shape model or use geometric shape features to represent 3D objects in a high-dimensional feature plane. The drawback of 3D descriptors is, that they are extremely memory intensive, and therefore the recognition speed becomes a significant challenge. The proposed novel 3D object search technique based on 3D object recognition reads the 3D vertices of a CAD file, down samples the triangular meshes, further finds the center of each triangle and then maps the points onto two 2D planes, namely the top view of the XY plane and the front view of the YZ plane. The points in each plane are clustered into ten clusters, and the cluster centres are used as the feature descriptor for the 3D object. The proposed technique improves the search time with speed and classification accuracy significantly over existing shape descriptor-based techniques.