Low-Power Image Recognition Challenge (LPIRC 2018)
NEW! IEEE Announces 2018 LPIRC-II
Due to the large number of contestants and strong interests of sponsors, the IEEE International Low-Power Image Recognition Challenge (LPIRC) will be held again in November 2018, called LPIRC-II.
LPIRC-II offers two tracks:
Track 1: Evaluates accuracy and execution time using TensorFlow models
Track 2: Evaluates energy consumption using the Caffe2 deep learning framework (software) running on NVIDIA Jetson TX2 (hardware)
More details of these two tracks are available after registration at lpirc.ecn.purdue.edu. All contestants solutions must be submitted online.
Submissions will be open 1-15 November. Each team can submit at most twice every 24 hours. The winners will be decided on 20 November and the winners will be notified.
Please notice that due to the Thanksgiving holidays in USA, the submission deadline will not be extended.
Prizes: First prize in each track: $1,500, Second prize in each track: $500. The winners will be invited to speak in a major conference early 2019.
To win a prize, a team's solution must obtain scores higher than 2018 LPIRC-I champions' scores. LPIRC-I was held in May-June 2018.
Track 1: 0.7267
Track 2: 0.285212 (mAP: 0.389811, energy: 1.36674 HW)
LPIRC 2018 was held June 18 in Salt Lake City, Utah
Download IEEE LPIRC 2018 Press Release (PDF, 551 KB)
LPIRC organizers and Track 1 winners.
LPIRC organizers and winners of Tracks 2 and 3.
The 2018 IEEE International Low-Power Image Recognition Challenge (LPIRC) has successfully concluded on June 18 in Salt Lake City, co-located with the IEEE Conference on Conference on Computer Vision and Pattern Recognition (CVPR). This is the fourth LPIRC; 21 teams competed in three different Tracks. In total, the teams submitted 131 solutions. LPIRC is the only competition that evaluates computer vision technologies by accuracy, execution time, and energy consumption together. Each team must develop a solution that can identify objects (such as humans, cars, tables) in images and mark their locations in the images. The first track, a new track sponsored by Google, evaluates accuracy and execution time. The second track, sponsored by Facebook, uses the Caffe2 deep learning framework running on NVIDIA Jetson TX2. The third track, unchanged from the first LPIRC in 2015, has no restriction in software or hardware.
This year’s winners are
- Track 1: Qualcomm
- Track 2: Seoul National University
- Track 3: ETRI and KPST
The top score of Track 2 is more than twice of the 2017 top score. The top score of Track 3 is nearly four times of the 2017 top score. Since 2015, the score has improved by 24 times. “The purpose of LPIRC is to identify the state-of-the-art in computer vision. Thus, it is important to see significant improvement year after year.”, said Terence Martinez, Program Director, Future Directions of IEEE Technical Activities. LPIRC started in 2015 as part of the IEEE Rebooting Computing Initiative, co-chaired by Elie Track and Tom Conte.
LPIRC uses ImageNet as the training data. The referee of LPIRC is open-source and contestants can replicate the competition environments. On June 18, the researchers from Facebook and Google explained how to apply computer vision on mobile systems. More than 100 people attended the presentations. “We will likely see sophisticated vision technologies on mobile phones in the coming years.”, said Fei Sun from Facebook. Bo Chen from Google, a member in the organizing committee, said, “This competition has created an infrastructure to evaluate the energy efficiency of vision technologies. We expect to see acceleration of improvements.”. Jaeyoun Kim, also from Google, said, “LPIRC is unique because participants need to make their entire systems work.”
Alex Berg, a professor from University of North Carolina, said “Over the past four years accuracy improved 13 times but the energy consumption has not reduced much.” Another member in the organizing committee, Professor Yiran Chen from Duke University said, “We expect that future winners would need to adopt innovative hardware platforms for better energy efficiency. There is still a lot of room for improvements.” Yung-Hsiang Lu from Purdue University thinks future mobile systems could do much more than detecting objects in image, such as understanding behavior and intention in video.
Immediately after the conclusion of 2018 LPIRC, the organizing team has started planning the next competition. After four years of successful LPIRC, the team considers to adopt more challenging tasks using new sets of data.
Attendees listen to presentations about creating energy-efficient computer vision technologies.
Thanks to all participants, organizing committee members, and student assistants.
Congratulations to LPIRC 2018 Track 1 First Prize Winner:
Congratulations to LPIRC 2018 Track 2 First Prize Winner:
Donghyun Kang, Seoul National University
Congratulations to LPIRC 2018 Track 3 First Prize Winners:
ETRI, KPST, South Korea
|LPIRC 2018 Organizing Committee|
|Alex Berg (UNC)|
|Achille Brighton (Google)|
|Yiran Chen (Duke)|
|Bo Chen (Google)|
|Andrew Howard (Goolge)|
|Jaeyoun Kim (Google)|
|Yang Lu (Facebook)|
|Yung-Hsiang Lu (Purdue)|
|Terence Martinez (IEEE)|
|Fei Sun (Facebook)|
|Matthew Ardi (Purdue)|
|Hsin-Pai Cheng (Duke)|
|Wenzhong Duan (Purdue)|
|Bo Fu (Purdue)|
|Hanwen Huang (Purdue)|
|Xin Liu (Duke)|
|Deeptanshu Malik (Purdue)|
|Eunbyung Park (UNC)|
|Shijin Wang (Purdue)|
|Jingchi Zhang (Duke)|
Financial Sponsors: IEEE Rebooting Computing, Google, and Facebook.
Technical Sponsors: Google and Facebook.
For inquiries, please contact firstname.lastname@example.org.
LPIRC 2018 Gallery
Many mobile systems (smartphones, electronic glass, autonomous robots) can capture images. These systems use batteries and energy conservation is essential. This challenge aims to discover the best technology in both image recognition and energy conservation. Winners will be evaluated based on both high recognition accuracy and low power usage.
Image recognition involves many tasks. This challenge focuses on object detection, a basic routine in many recognition approaches. The following two examples illustrate the task. In the first example, there are two objects: a bird and a frog. In the second example, there are several objects: cars, persons, motorcycle, and a helmet. The training and validation data for LPIRC comes from the ImageNet Large Scale Visual Recognition Challenge detection competition. The test data will be specific to LPIRC.
Retrospect and Prospect of LPIRC (PDF, 2 MB), presented at DATE (Design and Test in Europe), March 2018
Three Years of LPIRC (PDF, 3 MB), presented at ASPLOS, March 2018
Previous Competition Websites: