Quasi Spin Images.

The increasing adoption of 3D capturing equipment, now also found in mobile devices, means that 3D content is increasingly prevalent. Common operations on such data, including 3D object recognition and retrieval, are based on the measurement of similarity between 3D objects. A common way to measure object similarity is through local shape descriptors, which aim to do part-to-part matching by describing portions of an object's shape. The Spin Image is one of the local descriptors most suitable for use in scenes with high degrees of clutter and occlusion but its practical use has been hampered by high computational demands. The rise in processing power of the GPU represents an opportunity to significantly improve the generation and comparison performance of descriptors, such as the Spin Image, thereby increasing the practical applicability of methods making use of it. In this paper we introduce a GPU-based Quasi Spin Image (QSI) algorithm, a variation of the original Spin Image, and show that a speedup of an order of magnitude relative to a reference CPU implementation can be achieved in terms of the image generation rate. In addition, the QSI is noise free, can be computed consistently, and a preliminary evaluation shows it correlates well relative to the original Spin Image.