Page 19 - EETimes Europe June 2021

EE|Times EUROPE 19

OPINION | EMBEDDED VISION

Five Trends to Watch

in Embedded Vision

and Edge AI

By Jeff Bier, BDTI

While deep learning remains a dominant force, deep
neural networks alone don’t make a product.

PRESENTED AS A VIRTUAL EVENT in May, the Embedded
Vision Summit (bit.ly/3tXfI1J) examined the latest developments
in practical computer vision and AI edge processing. In my role as
the summit’s general chair, I reviewed more than 300 great session
proposals for the conference. Here are the trends I’m seeing in the
embedded-vision space.

DEEP-LEARNING DOMINANCE
First, surprising no one, deep learning continues to be a dominant
force in the field. It has radically changed what’s possible with computer vision. It has made
development more data-driven than code-driven, and it’s changed the tools and techniques we
use. But data is a pain. Where do you get it? How much of it do you need? How do you get more
of it? How do you know you have the right kind of data?

COMPLEX VISION PIPELINES
Second, despite the deep-learning revolution, product developers are increasingly realizing that
deep neural networks (DNNs) do not, by themselves, constitute a product. Real-world products
require a complex vision pipeline, often including camera and image processing, DSP, Kalman PROCESSORS APLENTY
filters, classical computer vision, and maybe even multiple DNNs, all combined in just the right Fifth is, honestly, an embarrassment of
way to get the results you need. processor riches. A year or two ago, I observed
that we were in a Cambrian explosion of pro-
DEMOCRATIZED DEVELOPMENT cessors for AI. Today, if anything, that trend
The third trend is democratization. It’s easier than ever to develop an embedded-vision appli- has accelerated and spread: It seems like
cation; thanks to a proliferation of tools and libraries, you don’t have to develop your algorithm everybody who makes a processor — whether
from scratch in assembly or C. A great example of this is Edge Impulse, which offers easy-to-use it’s a one-dollar MCU or a big, multicore,
software tools that let developers quickly and easily develop AI models and deploy them on low- multi-gigahertz, on-premises server pro-
cost microprocessors — all with very little coding required. cessor — is targeting edge-AI and vision
Also, we’re starting to see suppliers stepping up to support the whole pipeline (Lattice and applications.
Qualcomm are good examples here). It’s not hard to imagine a future in which a semiconductor That said, it’s a big space, and processor
company that has great tools for one component of the pipeline — DNNs, for example — but companies often target different zones in
nothing for the other critical pieces will lose market share to competitors that offer more com- terms of performance, price, and power. For
plete solutions. system developers, while it’s great having
a choice, it can be challenging to choose,
RISE OF PRACTICAL SYSTEMS especially when you consider not just tech-
Fourth is what I’d call the maturation of the field: We’re moving past the “wow, that’s so cool” nical factors (such as performance and power
stage and are asking how we deploy this technology in ways that are commercially viable and consumption) but other critical issues, such as
maintainable. price, business, and supply-chain risk.
Containerization is a great example. The approach has been a best practice in cloud devel- If there’s a megatrend here, it’s this: We’re
opment for over a decade, but we’re starting to see it used to speed development in practical living in a golden era of innovation in embed-
embedded systems, including vision and AI systems (which bring their own challenges, with ded vision. There’s never been a better time to
potentially frequent over-the-air model updates). build vision-based products. ■
Similarly, the specters of security and privacy rear their heads. How do we design systems that IMAGE: SHUTTERSTOCK
are secure against hackers and protect user privacy? Relatedly, how do we meet functional safety Jeff Bier is president of consulting firm BDTI,
requirements — indeed, how do we even test for such things? These are issues that don’t come founder of the Edge AI and Vision Alliance, and
up in science fair projects but do arise when you’re shipping real products to serious customers. general chair of the Embedded Vision Summit.

www.eetimes.eu | JUNE 2021