A Highly Optimized GPU Batched Elasticnet Solver (BENS) with Application to Real- Time Keypoint Detection for Image Retrieval

In this paper, we present a highly optimized GPU batched elastic-net solver (BENS) with application to real-time key-point detection for image retrieval. BENS was optimized to perform hundreds of thousands of small elastic-net fits by batching each fit from specific steps in the elastic-net computation into a large matrix multiplication which can be computed efficiently using the CUBLAS library. The main motivation for BENS was a real-time implementation of the Sparse-Coding Key-point detector (SCK) algorithm which has reaching applications in science, engineering, social science and medicine. When BENS was applied to accelerate SCK, we have achieved a 232x speed up compared to the original CPU implementation of SCK. To demonstrate the newly accelerated SCK algorithm, we conducted an Bo Vw based image retrieval experiment using SCK as the key-point detector.