During the three years of LPIRC, participants have
brought a wide range of platforms to the competition, including mobile phone, tablet, laptop, desktop, field programmable gate array (FPGA), and
experimental boards. The champions of the first
three years all used experimental boards with graphics processing units (GPUs). Recognition accuracy is
calculated based on the entire set of test data; thus, a
winning solution must be able to recognize objects
in many images. The 2015 winner used Fast Regions
with Convolutional Neural Network (Fast R-CNN) as
the foundation running on NVIDIA TK1. The 2016
winner used a convolutional neural network running, together with BING and FAST, on NVIDIA Jetson TX1. The 2017 winner used Tiny YOLO on Jetson
TX2. Articles canvassing the LPIRC experience and
the solutions of each year’s winner are listed in table
2. LPIRC’s sponsors include IEEE Rebooting Computing, IEEE GreenICT, the IEEE Council on Superconductivity, the IEEE Council on Electronic Design
Automation, NVIDIA, and Xilinx. In 2018, the sponsors included Facebook and Google.
In the first two years, LPIRC had a second track
that allowed offloading. A participant’s system has
two parts: a front end that communicates with the
referee system and a back end (another computer or
one or more cloud servers) for the computation.
Only the power of the front end was measured. In
2015, only one team entered the offloading track,
and that team was unable to successfully recognize
any image. In 2016, no team entered this track. In
2017, this track was no longer offered. In 2016, a
third track was added: the images were acquired by a
camera, rather than as files through the network. The
purpose was to use the camera to simulate the
human eye. Only one team entered this track and the
score was substantially lower than the same team’s
solution using the network. In 2017, this track, too,
was abandoned. In 2017, only one track was offered:
images were acquired through the network and
offloading was disallowed. The scores reported in
table 1 are from this main track since it was constant
over the three years.
In 2018, LPIRC will include two different tracks.
The first track keeps the same rules as the first three
LPIRCs: the competition is on-site at CVPR and each
team can bring any hardware or software platform.
The second track will provide a software development kit for a chosen platform. In this second track,
participants submit their solutions before CVPR and
their solutions are evaluated in this predetermined
1. More details about LPIRC can be found at rebootingcom-puting.ieee.org/lpirc
Yung-Hsiang Lu is a professor in the School of Electrical
and Computer Engineering and (by courtesy) the Department of Computer Science of Purdue University. He is an
ACM distinguished scientist, an ACM distinguished speaker, and a member in the organizing committee of the IEEE
Rebooting Computing Initiative. Lu is the lead organizer of
the Low-Power Image Recognition Challenge, and the chair
of the Multimedia Communication Systems Interest Group
in the IEEE Multimedia Communications Technical Committee.
Alexander C. Berg is an associate professor in the Department of Computer Science at the University of North Carolina at Chapel Hill. His research concerns computational
visual recognition. He has worked on general object recognition in images, action recognition in video, human pose
identification in images, image parsing, face recognition,
image search, and machine learning for computer vision.
He co-organizes the ImageNet Large-Scale Visual Recognition Challenge.
Yiran Chen is an associate professor at Duke University and
serves as the codirector of the Duke Center of Evolutionary
Intelligence. His research focuses on emerging memory and
storage systems, deep learning acceleration and neuromor-phic computing, and mobile computing. He is a recipient of
an NSF CAREER award and the Outstanding New Faculty
Award from the ACM Special Interest Group on Design
Table 1. Score Improvement, 2015–2017.
Ratio Score mAP Year Number of Solutions Energy (WH)
Table 2. Articles Detailing the LPIRC Experience and Winners’ Solutions.
2017 CNN-Based Object Detection Solutions for Embedded Heterogeneous Multicore SoCs. Asia and South
Paci;c Design Automation Conference.
2017 Low-Power Image Recognition Challenge. Asia and South Paci;c Design Automation Conference.
2016 Towards Real-Time Object Detection on Embedded Systems. IEEE Transactions on Emerging Topics in
2015 Rebooting Computing and Low-Power Image Recognition Challenge. IEEE/ACM International
Conference on Computer-Aided Design.