Notice

April 19, 2022 Winners' reports and implementation source code have been released.

April 7, 2022 The video recording of the awards ceremony is now available on YouTube. You can also view it at the bottom of this page.

March 16, 2022 Winners have been announced. Thank you very much to everyone who submitted the final result set.

February 21, 2022 The dataset has been made available again due to the extended deadline for submissions. The leaderboard for object tracking evaluation for the extended period is available.

February 18, 2022 The deadline for submissions has been extended to Wednesday, March 2, 2022.

February 7, 2022 The video of the online seminar has been released.

February 3, 2022 The Final Submission Application Form has been released.

January 27, 2022 The documents of The 2nd Online seminar has been released.

January 18, 2022 The MOTA threshold was fixed at 0.6.

December 22, 2021 The documents of The RISC-V reference implementation has been updated.

December 17, 2021 We revised the provided board accessories. Provided board includes an AC adapter, and power conversion cable.

December 6, 2021 We are currently accepting applications for participation in the online seminar on The RISC-V reference implementation. Click here to apply!!

December 3, 2021 The RISC-V reference implementation is now available on Google Drive. Also, a modified version of Vivado.zip has been uploaded.  If you have already downloaded Vivado.zip, we apologize for the inconvenience, but please download it again.

November 30, 2021 RISC-V reference environment has been released.

With the progress of artificial intelligence (AI) technology, social implementation such as image recognition using AI technology, automatic driving, and natural language processing is rapidly progressing. Especially in the edge computing field, since it is necessary to realize AI technology with higher efficiency, startups mainly in the United States and China and major vendors are accelerating their entry into AI hardware, and even in Japan Not only that, there is an urgent need to develop human resources and industries for hardware to accelerate AI processing such as LSI and FPGA.
Against this background, we will carry out this contest, in addition to the conventional AI technology development centered on software, we are aiming to develop human resources and startups with AI hardware in mind, as well as industry development utilizing these technologies. In this contest, we will develop AI hardware equipped with the "RISC-V" chip, which is currently attracting attention, and set the hardware system development that considers edge computing including hardware and software (network model and system optimization) as an issue.

The 1st AI Edge Contest (Algorithm Contest (1))
The 2nd AI Edge Contest (Implementation Contest (1))
The 3rd AI Edge Contest (Algorithm Contest (2))
The 4th AI Edge Contest (Implementation Contest (2))

*Please refer to the winners' solutions in the past AI edge contest.

Contest Outline

Subject (Algorithm development)
Allocate a rectangular area containing objects to be predicted as a bounding box = (x1, y1, x2, y2) to the video captured by the vehicle front camera, and give any unique object ID to the same object in each video. With allocations of multiple bounding boxes to each frame in the videos, the bounding box is represented by specifying four coordinates with the upper left corner as the origin (0,0), upper left coordinate of the object area (x1, y1), and the lower right coordinate (x2, y2).

However, the object to be evaluated (objects to be inferred) is limited to those that satisfy all of the following.
・There are two categories: “Car” and “Pedestrian”
・Objects with 3 or more frames in each video (Frames do not have to be continuous)
・Objects with a rectangle size of 1024 pix² or more

*Even if an object appears in three or more frames in the video, if there are less than 1024 pix² or more in three frames, the object including the size of 1024 pix² or more will not be evaluated.
*Please note that if an object that does not meet the conditions is detected, it is considered as a false detection.

Subject (Algorithm Implementation)

Implement the developed algorithm on the target platform equipped with a RISC-V chip.

The section to be optimized and the processing time to be measured is defined as follows.
Time to perform inference processing on the video data loaded in the memory and then convert it to JSON-equivalent information such as the ID and rectangle information of each object.

Carefully check the RISC-V usage requirements described in the "Rules" tab before implementing.

Data

Travel Route

Flow of entry / registration for the contest
1. Sign up for SIGNATE (Please register through the member registration button on the top right of the page if you are not currently a member.).
2. Download the data from the data tab (Please agree the terms of entry that will be displayed when downloading.).
3. Develop an algorithm and implement / optimize it on the selected target environment
   * We plan to provide FPGA board to a maximum of 50 applicants based on the review of application among the people who apply through application form for the board offer document review. (The number of boards offered may increase.)
   * If you do not have an Ultra96-V2 board and would like to use it as a target environment, please purchase it yourself or apply by clicking the button below.

Application Form for Board Offer Document Review

4. During the contest, submit the object trakking result from the "Submit" button on the upper right of this page and check the MOTA score on the leaderboard.
   * Please refer to the MOTA function that can be downloaded on the data page.
   * The MOTA score until the last day of the contest (public score) and the MOTA score after the contest (private score) are the same.

5. (Optional) During the contest, we will submit the provisional processing time ranking as reference information, so please cooperate.
   * Please submit the average processing time [ms] per video at the time of posting and the evaluation result by MOTA from the "Submit" button on the page that transitions from the following.
   * Submits are accepted at any time until the deadline of 2022/1/15, and can be updated.(The latest results will be displayed)
   * In the case of a team, the team leader should submit it as a representative. The user name of the poster is displayed on the interim processing time leaderboard.

Provisional processing time submission form

6. Submit the final result set by the end of the contest.
   * For details, refer to "submission_details.pdf" that can be downloaded from the data page.
   * Upload on the Internet such as Dropbox, Gigafile, etc., and inform us of the information required for downloading to the following form.

Final Submission Application Form

ONLINE Award Ceremony

The AI ​​Edge Contest is also collaborating with the 3rd Japan Automotive AI Challenge sponsored by the Society of Automotive Engineers of Japan to promote the discovery and training of engineers.

Decision method of winners
The secretariat will confirm the submitted final result set, and from those who meet all of the following requirements, they will decide the prize winner based on the processing performance time (ms) for the processing speed award and the content of the report for the idea award. An online interview may be set when necessary, centering on the idea prize, and a presentation (assuming 5 minutes presentation + 10 minutes Q&A) will be held such as implementation ideas.

Evaluation of Recognition Accuracy
・ It is evaluated based on the evaluation function “MOTA (multiple object tracking accuracy)”
・ The maximum evaluation value is 1, and the higher the accuracy, the larger the value. It can be negative.
・ Please refer to the implementation example of the evaluation function which can be downloaded from the data page.

1. When the overlap (IoU) of the prediction area and the correct area (rectangle) exceeds 0.5 for all the frames in an video, it is determined that the area can be detected.
2. Calculate MOTA for each subjected prediction category based on the detected result.
3. Calculate the average of 2. of the all subjected prediction categories.
4. Calculate the average of 3. of the all videos for evaluation.

Reference: https://cvhci.anthropomatik.kit.edu/~stiefel/papers/ECCV2006WorkshopCameraReady.pdf

About implementation report

It is not necessary to cover everything, but if you appeal to these points, the points for judging the idea award will be high.
・Outline of implementation and demo operation
    - System configuration/architecture outline of implemented hardware in FPGA
    - Operating frequency, performance, etc.

・Content of ingenuity and appeal points (Especially from the viewpoint of RISC-V utilization)
    - Ingenuity on algorithm (network model used, learning method such as assembly, improvement of data set, etc.)
    - Architectural innovations (optimization of speed and area)
    - Electrical consideration (measurement comparison with wattmeter etc.)
    - Others (cost considerations, devices for shortening development period, etc.)

・The report itself
    - Easy to understand each issue, clarification of ingenuity, etc.

Premise of participation
・ Please understand the purpose set in the competition and build a system that is conscious of practicality. Practicality here means that it is fully automatic processing, high accuracy, high speed, low calculation amount (HW cost saving), scalable with respect to the amount of data, and high model interpretability.
・ Participation by cheating or ignoring the rules is not allowed.
・ It is prohibited to slander others or to act against public order and morals. If the secretariat determines that it is malicious, please be aware that there is a possibility of deprivation of winning qualifications and membership qualifications before participating.

Use of System
・ You may have one account per participant.
・ If you want to join as a team, you can create a team of up to five people by Jan 31, 2022. (Click here for how to create a team)

Handling of Information
・ Public disclosure of algorithms, ideas or other materials created to participate in this competition shall not be restricted, only if it is clearly stated that those are created to participate in this competition at the time of publication.
・ However, at the public disclosure, please specify it (including links) in the forum of this competition so that all participants can view it.

Use of Data
・ Use of the data other than provided data, and learned data shall not be restricted if the method allows verification of reproducibility for free. For example, the dataset with non-profit purpose, such as Waymo Open Dataset, can be used.
・ However, use of data provided by past AI Edge Contests is prohibited.
・ Modification of training data (including manual labeling and label rewriting) is allowed.
・ For API, the data that can be republished as an open source can be used.
・ Inference is performed on video-by-video basis. When tracking an object in videos, future information can be used if it is the data for test in the same video (tracking two-way). However use of the data in other videos is prohibited.

Creating algorithms
・ If you enter new data in the format provided in the competition, only the model for which the prediction result is automatically output will be evaluated.
・ The proposed method should be reproducible without any additional costs (you are not allowed to not use any paid external APIs, etc.), and the algorithms are meant to be used after this competition.

Way to implement
Use RISC-V Cores during object tracking tasks.
The cases where RISC-V is used in this contest are permitted and not permitted below.

【Permitted cases】
・ RISC-V is used as microcomputer controller during Deep Neural Network processing on FPGA.
・ Participant designed some special operator by extending RISC-V ISA to accelerate Deep Neural Network processing on FPGA
・ Vitis AI is used for object tracking, and after that, RISC-V is used in the processing on the video data loaded in the memory and then convert it to JSON-equivalent information such as the ID and rectangle information of each object
・ (For Idea award) Use the SoC K210 equipped with RISC-V, and implement object tracking process with Deep Neural Network core embedded in SoC

【NOT permitted cases】
・ Use RISC-V as result output only (not used in core processing or controller of component potion.)
・ Use a emulator of RISC-V (e.g. QEMU) on the top of ARM, and process the core part  in the emulator

The purpose of this contest is to develop human resources involved in the technology and application of the RISC-V, which is expected to be used in the Edge AI field in the future, and to educate the public on how to use RISC-V.

Judging criteria may be added in the future at the discretion of the Steering Committee or other relevant committee members at least one month before the contest deadline. If other cases arise, the review committee will deliberate on them and make a decision on an individual basis.

Host：Ministry of Economy, Trade and Industry / New Energy and Industrial Technology Development Organization (NEDO)

    

Co-host：SIGNATE Inc.
Support：IEEE Japan Office / The Society of Instrument and Control Engineers (SICE)
Sponsors：Society of Automotive Engineers of Japan / Google Japan LLC / Sony Corporation / SoftBank Corp. / Sompo Japan Nipponkoa Insurance Inc. / Digital Media Professionals Inc. / Panasonic Corporation / NSITEXE Inc. / Tier IV, Inc.

Endorsement：IEEE Japan Office / The Society of Instrument and Control Engineers / The Japanese Society for Artificial Intelligence / The Institute of Electronics, Information and Communication Engineers / The Japan Society for Precision Engineering / Japan Electronics and Information Technology Industries Association / The Robotics Society of Japan

Steering Committee Members

Reference information for obtaining the Ultra96 board to use for this contest

・Subjected board for this contest is Ultra96-V2.
・Regarding method to obtain: There are two ways to obtain it.

1. Provided from this contest
- Although limited in number, the Ultra96 board will be provided after review for the contest participants. For those of you who wish to be provided, please apply through the "application form for board offer document review".
- Provided board includes an AC adapter, and power conversion cable. and JTAG board.

2. Obtain from the Avnet website
- You can purchase from here.
- We recommend you to purchase the following related parts along the purchasing of the board.
・JTAG board for Ultra96: It is necessary to connect the UART (serial) console.
・Avnet recommended AC adapter, power conversion cable
　- Power supply
　- Conversion connector

Regarding the errata of Ultra96 board
The errata from Avnet is published the following site.
http://avnet.me/Ultra96-V2_Errata

The errata particularly has the following known issues:
・Reboot issue of Ultra96-V2
　- Ultra96-V2 When the load of the PL part goes up high, MPSoC reboots.
　- Ultra96-V2 It is posted on the errata reporting page.
This is caused by the reset operation in MPSoC that was triggered by the PowerGood signal's drop when a specific threshold is exceeded while the IC called PMIC is monitoring the Vccint voltage.
- In the current V2 board, a reset is applied when a set threshold is exceeded due to increase in current consumption of the PL section.
- How to distinguish the PL power supply current setting is as follows.

・Also for boards with power limits of 3.5A and 4.0A, The limit of this errata can be relieved to about 5.5A by revising the board and changing the setting of PMIC. A complete set of countermeasure procedures prepared by Avnet can be downloaded from below.

　20191114_Ultra96-V2 PMIC update procedure.zip (This can be obtained by the data tab on this website)
　Infinion that is required for the measures of this issue.
　The sales information of USB005 writing cable is as follows.
　Mouser
　Digikey
　Chip1Stop

・We plan to provide a board that supports 5.5 A of PL power supply current for the board offered from this contest.

Mr. Nakahara, a member of the Steering Committee, created a tutorial for beginners when applying for this contest.(Japanese only)

You will license your winning Submission and the source code used to generate the Submission under an Open Source Initiative-approved license (see www.opensource.org) that in no event limits commercial use of such code or model containing or depending on such code. As a reference, we will post a comparison table of representative open source writers.

Q. There was an obvious mistake in the label of the training data. What should I do?
A. It is possible to correct the learning data according to the rules, so please delete or correct it. We check all kinds of data, but it is difficult to detect and correct 100% of mistakes, so we take the present result as "correct" and analyze it. Thank you for your understanding.

Q. What happens to the ID when the same object is actually less than 1024 pix² in size, hidden behind another object, or left off the screen and then returned? 
A. Even if they are discontinuous, if they are actually the same object, they will be given the same ID.

Q. If there are the same objects in frames 1 to 3 at 2000, 2000 and 1000 pix² respectively, will only the objects in frames 1 and 2 be judged as labels? Or, since the number of frames is less than 3 as a result of being filtered by the rectangle size condition earlier, is it not all included in the label as a result?
A. All are not included in the correct answer label because they are filtered by the condition first. Even if an object appears in three or more frames in the video, if there are less than 1024 pix² or more in three frames, the object itself, including 1024 pix² or more, will not be evaluated.

Q. It seems that the code for calculating MOTA that can be downloaded on the data page does not include filtering of the three conditions to be inferred such as category, number of frames to appear, and size of rectangle.
A. When actually using the distributed MOTA code and verifying it in a local environment, please apply the conditions to the annotation that is the true value. If an object that does not meet the conditions is detected, it will be treated as a false detection. Therefore, when creating a post file, please post by yourself with the contents excluding the objects that satisfy the conditions removed.

Q. For "objects with more than 3 frames", is the first 2 frames that appear also considered in the judgment?
A. Yes. The judgment is made from the first frame.

Q. Do "objects with more than 3 frames" need to be continuous?
A. No. It does not have to be continuous.

Q. I implemented it on an Ultra96 board. Please tell me how to make an SD card image for submission.
A. Please refer to the forum / article of the 2nd contest below for how to create an SD card image.

https://signate.jp/competitions/191/discussions/sd

▼ Special Contract on this Contest
In accordance with Article III “Prizes and Rights” of the special regulations established for this contest as described below, participants will either choose to make their own submission data, source code and other relevant items be licensed under an Open Source Initiative approved license (www.opensource.org) or not. This license imposes no restrictions on commercial use for licensed models, source code, and any results attained using said source code. Incidentally, the participants who will not choose this license will not be able to prize winner candidate, even though they can participate in the contest. In regard to posting and/or other publication on any social media site, blog or similar by the participant of information relating to any algorithm developed by said participant or any other materials produced for or related to this contest (including materials used for final submissions and/or final evaluations/inspections), no restrictions apply to posting and/or publication as long as the participant clearly explains that the posted/published items relate to this contest, and absolutely posts the positing and/or publication information (including the link URL to the pages) on the contest forum in order to let all the participants know it.