OpenVL is a high-level task-based abstraction for computer vision which does not require extensive knowledge or experience with vision methods, unlike most frameworks which present APIs as lists of specific techniques. OpenVL requires developers to have enough knowledge of a task to accurately describe it using our API; the description is analyzed and an appropriate method is invoked to provide a solution. We present our methodology for accelerating OpenVL on heterogeneous platforms using OpenCL for fundamental operations such as segmentation and correspondence. The accelerated methods are combined with CPU-only to offer greater functionality; due to the effectiveness of the abstraction, all of the methods are hidden to the developer, leading to an efficient and mainstream-developer friendly computer vision API. An evaluation on AMD OpenCL-compatible APU and GPU is presented to demonstrate the advantage of an abstraction which provides mainstream developers with performance gain and energy reduction by utilizing the resources efficiently.

Presented in San Jose, November 2013 at the AMD Developer Summit.