2017 Embedded Deep Learning Design Contest
October 20, 2017
http://cmalab.snu.ac.kr/edldc
Call for Participation
Embedded deep learning design contest (EDLDC) is a competition for the energy-efficient design of a deep learning application under a given set of constraints. The first EDLDC will be held in conjunction with ESWeek 2017 as a part of the program of International Workshop on Highly Efficient Neural Networks Design (HENND). This year’s contest targets object detection which is one of the most popular deep learning applications on embedded systems, e.g., self-driving car, AI secretary, robot, surveillance, etc.
Register here for EDLDC 2017
Problem description
Given a set of images, the participant's design needs to find as many correct bounding boxes of objects as possible.
The evaluation metric is mean average precision (mAP), energy consumption, and the number of processed images during the given test period.
Tracks
Track A: Any hardware platform having power connector (110V or 220V), except NVIDIA TX2, can be used. The evaluation metric is mAP*(# images in 5min)/energy.
Track B: Only NVIDIA TX2 is used. The evaluation metric is mAP*(#images in 5min).
Contest rule
Training: The participants need to train or develop their designs at their sites with any training/validation data sets, e.g., Large Scale Visual Recognition Challenge detection competition. We use a private test set in the test server.
Test server: The participant's design needs to be connected to the test server via internet.
The design needs to receive N images from the test server, performs computation to obtain bounding boxes and class scores and returns the results to the test server.
The test server gives as output mAP and the number of processed images.
Energy measurement: The electric current is measure at the power outlet.
Thus, the design needs to be powered via the typical power outlet, not a battery.
Scoring
The scoring system will be the same as used in IEEE LPIRC. Download the test server and sample client code here.
Schedule
Schedule | Date |
---|---|
Call for participation and official website open for problem definition and registration | June 5, 2017 |
Registration deadline | September 1, 2017 |
On-site contest and award ceremony | October 20, 2017 (at HENND workshop, ESWeek 2017) |
On-site Contest
The on-site contest will be held at Peacock room (36F) of Lotte Hotel Seoul (ESWEEK 2017 conference venue), 9am-12pm, Oct. 20.
Prizes
Each track will have 1~3 winners. Certificates and cash prize will be given to the winners.
Note that we have minimum mAP constraint. Thus, participants who have mAP below the average of the participants will not be eligible for the prizes.
Winners
Track A
Winner | Participant Name | Affiliation |
---|---|---|
1st Prize | Jintaek Kang | Seoul National University |
2nd Prize | Yong-Deok Kim | Samsung Electronics |
Track B
Winner | Participant Name | Affiliation |
---|---|---|
1st Prize | Duseok Kang | Seoul National University |
2nd Prize | Donghyun Kang | Seoul National University |