SAN JOSE, Calif. — Developers of computer vision applications are expected to get a leg up from the OpenVX specification published today. OpenVX defines a high-level interface for mapping vision apps on to any chipset in a way optimized for low power.
The interface lets developers handle front-end processing of computer vision data without waking up a CPU or graphics core. “That enables low-power operation for virtual and augmented reality wearables that want to continuously scan the environment -- a very battery sensitive use case,” said Neil Trevett, president of Khronos Group, in an interview with EE Times.
Khronos is the ad hoc standards body behind OpenVX and other standards such as the popular OpenGL graphics API. The group will release an open-source implementation of OpenVX in C before the end of the year.
A broad range of companies announced support for OpenVX including AMD, Nvidia, Cadence, Imagination Technologies, Intel, and Samsung.
OpenVX aims to be complementary with the existing OpenCV computer vision API, which it positions as good for prototyping. However, OpenCV is less power efficient because it requires memory access for each operation and less portable across processors, Trevett said.
“OpenCV makes continual round trips to memory, which is the worst thing you can do, particularly for mobile.”
OpenCV also lacks backing of a consortium and conformance tests, which Khronos will develop for OpenVX, he added.
OpenVX breaks operations into steps that can be run serially or in parallel. It operates at a higher level of abstraction than OpenCL, an API geared for parallel programming multicore SoCs.
“A key challenge for engineers is efficiently mapping complex algorithms onto the processor best suited to the application,” Jeff Bier, founder of the Embedded Vision Alliance, said in a Khronos press statement. “OpenVX is an important step towards easing this challenge.”
— Rick Merritt, Silicon Valley Bureau Chief, EE Times