Parallelizing SIFT on a Distributed Memory Cluster using Data Decomposition
SESSION: Research Poster Reception
EVENT TYPE: Poster
TIME: 5:15PM - 7:00PM
AUTHOR(S):Stanislav Bobovych, Wesley Emeneker, Seth Warn, Amy Apon, Jack Cothren
ABSTRACT: Image registration corresponds the features of images that overlap. Scale Invariant Feature Transform (SIFT) can calculate image features. However, SIFT calculation for large satellite images is slow and requires expensive workstations with large memory.
In this project, SIFT is parallelized using both row and block decomposition on a distributed memory cluster. In one implementation, the master process reads the image and sends different portions of the image to different processes. Another implementation uses MPI IO.
All tested methods reduce SIFT runtime with linear speedup. Using block decomposition and in-memory partitioning, a 1.2 GB satellite image was SIFTed on 32 TACC Ranger nodes in 7.4 minutes. This processing was too large to perform before parallelization. The current implementation can register images with only minor data loss due to border effects.
Background about image registration and SIFT are presented, followed by description of the parallelization approach, results, and experiences running on Teragrid.