Multi-View 3D Geometry Reconstruction: Exploiting Massive Parallelism

3D geometric reconstruction from digital images captured from con-sumer cameras is an inexpensive, but computationally demanding ap-plication. In this experimental study, we have explored parallelism in the best known public domain software (Bundler and PMVS2) and found that massive parallelism exists at various levels that can be ex-ploited on various computer architectures (such as multi-cores, dis-tributed memory and GPU machines). Although in this report, we present results for multi-core and distributed memory machines, but we believe that similar results could be obtained on vector machines and GPU architectures as well. We have tested our MPI codes on a large clusters of 500 Linux nodes. Having shown almost linear scalabil-ity, we stress upon the need of improvements in sequential algorithms to enhance the quality of generated models. 1