PRCV2020 Challenge(2): 2nd High-speed and Low-power Visual Understanding Challenge

1. The purpose and significance of the competition

So far, artificial intelligence has achieved unprecedented achievements in the field of visual understanding, the performance indicators of various tasks have been constantly refreshed, and its results have been widely used in various fields of human life. However, in actual deployment, many systems and products have additional restrictions on algorithm speed and power consumption. With the expansion of the scale of artificial intelligence applications, such problems are increasingly prominent. At present, most researchers only focus on the performance improvement of visual algorithms. The consideration of speed and power consumption is still lacking, but it is obvious that in order to solve practical problems, algorithm power consumption, speed, and performance are important factors that need to be paid attention to. Therefore, in order to evaluate the current research status in the field of vision and build a general benchmark for algorithm comparison in three aspects: power consumption, speed, and performance, we will organize the "PRCV2020 The 2nd High Speed ​​Low Power Visual Understanding Challenge. The purpose of this competition is to further explore visual comprehension technology that takes into account visual comprehension performance and high efficiency and energy saving, on multiple platforms of CPU, GPU and small embedded AI chips, improve the research level of visual comprehension technology and promote the industrial development of computer vision discipline , To promote the application of visual intelligent analysis technology in practical problems.

2. Competition organizer

Organizers: Xin Zhao, Jing Zhang, Yan Bai
Organization unit: CASIA
Contact: 010-82544720
Email: yan.bai@ia.ac.cn
Address: No.95, Zhongguancun East Road, Haidian District, Beijing

3. Organization

The organizer tests the algorithm model submitted by the team on the virtual test environment (Ubuntu) by using the test data. By verifying the output results of the executable program, the performance, power consumption, and speed indicators of the model on multiple platforms are comprehensively evaluated. Ranking (in order to prevent cheating, the organizer does not publish the test set, but the test data is the same as the instance segmentation task in the Microsoft COCO dataset (non-removed categories, a total of 80 categories).

4. Schedule

  • June 1: The organizing committee publishes the competition tasks, way of registration and evaluation methods on the website.
  • June 1 - June 30: Registration and confirmation of qualifications.
  • July 1 - July 31: The competition starts, the organizing committee releases the competition verification set. Each team completes the competition task within the specified time, and uploads the result to the designated location according to the specified format. Each team can only submit the results once to participate in the evaluation (if you need to modify, please contact the organizing committee to make changes within the specified time).
  • August 1 - September 1: The judges verify the execution results of each team's procedures in accordance, evaluate the results following the prescribed evaluation method, compare and announce the ranking of each team.
  • October 16 - October 18: The competition unit introduced the competition methods of the winning units and awarded the prizes to the winning teams in the third session of China Pattern Recognition and Computer Vision Conference PRCV2020.

5. Other information

Visit the competition website for details: http://hislopvision.aitestunion.com/