Electronic Design
Q&A: Embedded Vision Alliance Founder Jeff Bier on Visual Intelligence’s Future

Q&A: Embedded Vision Alliance Founder Jeff Bier on Visual Intelligence’s Future

In this Q&A interview, Jeff Bier, founder of the Embedded Vision Alliance and co-founder of BDTI, shares why embedded vision is a technology to watch, and offers a preview of the fifth Embedded Vision Summit, May 12 in Santa Clara, Calif.

The Embedded Vision Summit is coming up soon, and a lot has been happening in the world of vision and 3D technology. I talked with Jeff Bier, founder of the Embedded Vision Alliance and co-founder of BDTI, about what might be at the Summit and some of the trends he’s observing these days.

Download this article in .PDF format
This file type includes high resolution graphics and schematics when applicable.

Wong:  This is your fifth Embedded Vision Summit. Tell us how the implementation of visual intelligence has progressed since the event’s inception.

Jeff Bier, Founder, Embedded Vision Alliance; Co-Founder

We’ve seen huge growth in activity among system developers who want to bring computer vision into their applications. Similar to how wireless communications has become pervasive in electronic systems over the past 10 to 15 years, embedded vision is beginning to be deployed widely across a broad range of systems. Many of the barriers to adoption have been overcome, and it’s now very practical to incorporate vision technology into systems. Enabling machines to see and analyze their environments is enabling a lot of innovative product capabilities and exciting new applications.

Wong: Speaking of applications, which ones have the most potential for embedded vision?

Bier: We see the proliferation of embedded vision into virtually every category of electronic products, from automobiles to drones, robots of many kinds, medical equipment, toys, and many more. For instance, Dyson recently introduced a robotic vacuum cleaner that uses 3D vision to map its path within a room. Mike Aldred of Dyson led the development of this innovative product, and is one of our keynote speakers at the Embedded Vision Summit this year. Mike will discuss opportunities, challenges, and techniques for bringing embedded vision to consumer products. Robert Chappell from EyeTech Digital Systems will address eye-tracking applications in medical, automotive, and headset applications.  And we have a number of talks on using deep neural networks for object recognition, which is a very hot topic. We’ve also added a conference track this year to explore market trends and the expanding business opportunities associated with embedded vision.

Wong: What embedded vision challenges that need to be solved by the industry will be addressed at the Summit?

Bier: One of the biggest challenges is that these computer vision applications demand a great deal of processor performance—typically tens of billions of operations per second—because they use complex algorithms operating on video data in real time. Getting that much processing power into a small, inexpensive, low-power device can be a challenge. And because algorithms are evolving rapidly, it’s important that processors be easy to program. Processor suppliers have responded with a variety of processors optimized for these applications. The Summit will include presentations on different types of heterogeneous processors for vision applications, and on a range of techniques for efficiently mapping vision algorithms onto these processors.

Wong:  What is the role of deep learning in enabling visual intelligence in embedded systems?

Bier: “Deep learning” refers to a class of neural-network algorithms that have recently been making headlines due to their ability to solve difficult problems in a number of domains—including vision problems like recognizing different types of objects. These neural networks represent a fundamentally different way of tackling vision problems. With traditional approaches, engineers develop complex procedural approaches to extract features of interest from an image, examine collections of related features, and reason about whether those features represent an object of interest.  In contrast, neural networks are fairly simple algorithms that learn to distinguish objects of interest through a training process during which they are shown large numbers of examples.

After decades of painstaking research, neural networks have recently been beating out traditional techniques on a number of difficult computer vision problems. As a result, companies are beginning to incorporate neural networks into their products. For example, the giant Chinese Internet company Baidu is using convolutional neural networks to enable image searches for users. Dr. Ren Wu from the Baidu Institute of Deep Learning will be discussing Baidu’s work in his keynote talk at the Embedded Vision Summit.

Wong:  Can deep learning be used for vision in embedded systems?

Bier:  Yes. Although deep neural networks are very computationally demanding, they are also fairly simple from an algorithmic point of view, and they have vast amounts of parallelism. This makes them well suited to specialized parallel processors, which can deliver a lot of performance per dollar and per watt. At the Embedded Vision Summit, we’ll be hearing from a number of innovators who are deploying clever techniques that make it practical to deploy deep neural networks in mobile phones, automotive safety systems, and other size-, power-, and cost-constrained applications.

Wong:  Since the debut of the Microsoft Kinect a few years ago, we’ve been hearing a lot about 3D vision.  Is 3D important?

Bier:  Yes. Just as 3D vision is very valuable for humans, it is very valuable for machines. And when the Kinect demonstrated that 3D could be a mainstream consumer technology, there has been a huge surge of investment and innovation in small, low-power, low-cost 3D vision capabilities. One of the fascinating things about 3D is that there are many ways to do it. Stereoscopic cameras, which work the same way our eyes do, represent the classic approach. But these days, there are also time-of-flight sensors, structured light sensors, structure-from-motion, and a variety of hybrid approaches. Presentations at the Embedded Vision Summit will cover several of these.

Wong: What else can developers expect to see and learn at the event?

Bier: The event is designed for product creators, innovators and business leaders interested in incorporating computer vision into their applications. It aims to deliver three things. First, ideas and inspiration about how vision technology can enable new and better products, such as cars that are safer today thanks to vision-based safety features. Second, through presentations, workshops, and demos, the Summit provides practical education for product creators about applications, markets, technologies, and techniques from expert practitioners. We cover applications, algorithms, processors, APIs, and development techniques.  And third, the Summit gives attendees opportunities to connect with experts, suppliers, and peers to share insights and get a leg up on their own product developments.

Download this article in .PDF format
This file type includes high resolution graphics and schematics when applicable.

References:

Embedded Vision Academy

Embedded Vision Alliance: Market Analysis

Embedded Vision Alliance: Video Interviews & Demos

Hide comments

Comments

  • Allowed HTML tags: <em> <strong> <blockquote> <br> <p>

Plain text

  • No HTML tags allowed.
  • Web page addresses and e-mail addresses turn into links automatically.
  • Lines and paragraphs break automatically.
Publish