Evaluating the block-parallel SIFT algorithm without boundary extension

The high computational complexity and great memory requirement prevent the SIFT algorithm from being processed realtimely. The block-parallel SIFT algorithm with boundary extension adopted by existing researches suffers from redundant storages or extra communications to process the boundaries of partitions. The block-parallel SIFT algorithm without boundary extension (pSIFT-noBE) can spontaneously overcome these disadvantages to gain a high speedup ratio at the cost of discarding a certain number of features existing in the boundaries. In this paper, the pSIFT-noBE is evaluated concerning the capacities of the feature extracting and speedup. Results from 30 test images in the USC-SIPI dataset show that, when the number of partitions are 4, 16, and 64, the pSIFT-noBE achieves 3.25X ~ 41.58X speedups over the serial implementation with 1.69% ~ 9.57% loss of extracted features. The pSIFT-noBE algorithm is applicable to the realtime processing of large-size images with rich features uniformly distributed.