GPU based geometric hashing for space partioning

Abstract

This paper presents a Graphics Processing Unit (GPU) based solution for classical geometric hashing and its variation with transformation functions on Speeded Up Robust Features (SURF) space of images. GPU based classical geometric hashing provides speed-up of 14.0x to 61.0x for offline indexing and 1.08x to 10.06x for online searching compared to sequential one for different partitioning sizes. GPU based transformation by mean invariancy and principal component based alignment with geometric hashing provides speed-up of 12.12x to 63.13x for offline indexing and 1.02x to 5.82x for online searching. This paper also proposes solution to execute multiple query simultaneously. It proves to be better than the serial execution of multiple queries. GPU based implementation of multi-query provide speed-up of 1.68x to 460.45x than the sequential one for online searching for multiple queries between 1 to 10 simultaneously. Experimentation is done using standard CASIA Palm-print based images.