Bookmark and Share

"Low-power Computer Vision: Status, Challenges and Opportunities," a Presentation from Purdue University

Register or sign in to access the Embedded Vision Academy's free technical training content.

The training materials provided by the Embedded Vision Academy are offered free of charge to everyone. All we ask in return is that you register, and tell us a little about yourself so that we can understand a bit about our audience. As detailed in our Privacy Policy, we will not share your registration information, nor contact you, except with your consent.

Registration is free and takes less than one minute. Click here to register, and get full access to the Embedded Vision Academy's unique technical training content.

If you've already registered, click here to sign in.

See a sample of this page's content below:


Professor Yung-Hsiang Lu of Purdue University presents the "Low-power Computer Vision: Status, Challenges and Opportunities" tutorial at the May 2019 Embedded Vision Summit.

Energy efficiency plays a crucial role in making computer vision successful in battery-powered systems, including drones, mobile phones and autonomous robots. Since 2015, IEEE has been organizing an annual competition on low-power computer vision to identify the most energy-efficient technologies for detecting objects in images. The scores are the ratio of accuracy and energy consumption. Over the four years, the winning solutions have improved the scores by a factor of 24.

In this presentation, Professor Lu describes this competition and summarizes the winning solutions, including quantization and accuracy-energy tradeoffs. Based on technology trends, he identifies the challenges and opportunities in enabling energy-efficient computer vision.

Omer
Offline
Last seen: 7 weeks 3 days ago
Level 1: Prestidigitator
Joined: 2018-07-19
Points: 1

This lecture is very important and significant

The most important conclusion was highlighted by the speaker, that currently GPU has the best performance, even when normalized to power. It will be interesting when new ASIC architectures like TPU, HAILO and Nervana will be mature enough for this competition.

Another important conclusion, is that under the metrics imposed by the competition, currectly, the real time low power accuracy is significantly lower than the best case unconstrained scenario. For reference, the best case scenario is comparable to the human ability.