Acceleration of the Pre-processing Stage of the MVS Workflow using Graphics Processors

Migrating CPU code to the CUDA programming language has been a challenge for some time. While the code for many high-performance and massively data-parallel applications has been successfully ported to GPUs, this task has received comparatively less attention for other applications that do not lend themselves so well to the characteristics of GPUs. Among the latter are applications of particular value to industry, where the use of GPUs can significantly improve the productivity of the system as a whole.

This article presents a real-world, industrial use case that shows (part of) a complex computer vision workflow. The processes in the workflow have low arithmetic intensity and work with simple data (integers). The main challenge is therefore to minimise the overhead of data transfers between the host and the GPU, or even within the memory of the device itself. While the speed-up achieved may not be as impressive as for applications that are perfectly tuned to the GPU architecture, the algorithms and data distribution proposed in this work allow a significant part of the overall workflow to be offloaded to the GPU, freeing the CPU to process other components of the workflow. As a result, this approach has a significant impact on the productivity and economic performance of the company.