Interview with Dr. Branislav Kisačanin, editor of the book Embedded Computer Vision

Computer Vision Central interviews Dr. Branislav Kisačanin, one of the editors of the book Embedded Computer Vision (Springer 2008)

1. Tell us a bit about yourself

My name is Branislav Kisačanin and since 2007 I've been with the Texas Instruments DSP R&D Center in Dallas. Before that I was with Delphi, a major automotive electronics supplier, for nine years. At Delphi, among other things, I contributed to the development of vision-based automotive safety algorithms and systems. I was one of the first, at least in the automotive domain, to work on embedded vision - the new discipline dealing with problems at the intersection of computer vision and embedded systems.

2. What sparked your interest in embedded vision?

One of my contributions at Delphi was the realization that due to the severe power and environmental limitations on automotive electronics, the automotive vision systems weren't going to be able to use the same processors as the PCs used to develop the algorithms. This is a big problem, because even if the algorithms execute in real time on a PC, there is no guarantee that they can achieve real-time operation on an automotive-grade embedded processor. This realization lead me to develop algorithms directly on an embedded processor, ensuring that in addition to yielding accurate results, the algorithms could execute in real time. One of the first such algorithms was the Fixed-Point PCA, a modification of popular eigenpictures (aka eigenfaces) for object detection and recognition. It achieves a nice speed-up by using a fixed-point representation of eigenvectors.

3. How did you get involved with the Embedded Computer Vision workshops?

It all started with a paper I presented in the first ECV Workshop (at 2005 CVPR in San Diego). I talked about some of the mathematical techniques I developed to speed up computer vision algorithms. It was a wonderful opportunity to hear about experiences of other researchers who, just like me, had to use limited resources to implement their vision algorithms. For example, I distinctly remember the presentation by Larry Matthies from JPL, who presented his team's work on vision algorithms involved in landing NASA’s rovers on Mars. At the time, along with the optical mouse and DARPA's Grand Challenges, it was one of the most visible examples of embedded vision.

At that time I was already involved in organization of conference workshops (RTV4HCI at 2004 CVPR, V4HCI at 2005 CVPR, and CVHCI at 2006 ECCV), and thought that ECV workshops are the next big thing to be involved in, as they dealt with really practical aspects of vision. As a result, I co-chaired the third and the fourth ECV workshops (at 2007 CVPR in Minneapolis and 2008 CVPR in Anchorage).

In addition to that, I also held embedded vision tutorials (at 2006 CVPR and 2007 ESC). I am currently preparing another one for the 2010 ICASSP.

4. Why is this book important?

The Embedded Computer Vision book combines invited chapters by Alan Lipton, Nik Gagvani, and others, with state-of-the-art chapters by researchers we met at the ECV workshop. It was co-edited with Sek Chai and Shuvra Bhattacharyya and judging by the fact that it has constantly been among the top-15 computer vision books on Amazon, the field of embedded vision is really a hot area of computer vision.

We see a continued growth in interest for embedded vision. Automotive vision already has a number of products in the market. To mention just a few, Volvo offers BLIS (Blind Spot Information System), which alerts drivers if a vehicle is detected in the driver’s blind spot, while BMW offers a night vision system (BMW Night Vision) capable of detecting people and animals even in pitch-dark situations and appropriately alerting the driver. You can see cool demonstration videos on YouTube. Automated surveillance (aka video analytics) is also taking off fast, with an ever growing number of companies offering systems for monitoring traffic, people, and property.

5. What advice would you give to people interested in entering the field of embedded vision?

First, I would encourage them to get involved in this exciting field. Second, I would quote Tom Huang from the Preface to the Real-Time Vision for Human-Computer Interaction.

Developing real-time robust HCI vision algorithms demands a great deal of "hack". The following statement has been attributed to our good friend Berthold Horn: "Elegant theories do not work; simple ideas do". Indeed, many very useful vision algorithms are pure hack. However, we think Berthold would agree that the ideal thing to happen is: An elegant theory leads to a very useful algorithm. It is nevertheless true that the path from elegant theory to useful algorithm is paved with much hack. It is our opinion that a useful (e.g., real-time robust HCI) algorithm is far superior to a useless theory (elegant or otherwise). We have been belaboring these points in order to emphasize to current and future students of computer vision that they should be prepared to do hack work and they had better like it.

This truth about computer vision is even more evident in embedded vision, but the rewards are great, as is often true at the intersection of two fairly removed disciplines, in this case computer vision and embedded systems.

Embedded Computer Vision by Branislav Kisačanin, Shuvra S. Bhattacharyya, Sek Chai

An interview with Sek Chai. Disclosure: Computer Vision Central is an official website for the book, Embedded Computer Vision.