CCF Big Data & Computing Intelligence Contest

The CCF Big Data&Computing Intelligence Competition (CCF BDCI) was founded by the Chinese Computer Society in 2013 and is a large-scale challenge competition for algorithms, applications, and systems in the field of big data and artificial intelligence. The competition is aimed at collecting demand from key industries and application fields, guided by cutting-edge technology and industry application issues, with the goal of promoting industry development and upgrading. Through a crowdsourcing and crowdsourcing approach, it gathers the wisdom of various industries, universities, and research at home and abroad, and has discovered and cultivated a large number of high-quality data talents for society.

The competition has been successfully held for ten times, attracting over 180000 participants from 25 countries, over 1500 universities, over 1800 enterprises and institutions, and over 80 research institutions worldwide. It has become one of the most influential, large-scale, and mature comprehensive competitions in the field of big data and artificial intelligence in China.

In 2023, we will welcome the 11th CCF BDCI, which has been striving for innovation for over a decade, continuously exploring new advantages in data value, and building a positive data ecosystem atmosphere. CCF BDCI will further expand its influence, focus on the development of digital economy technology and talent cultivation, and help promote the development of big data technology and industrial ecology in China.

• Background of the competition question

The Stable diffusion model is an image generation method based on the diffusion process, which can accurately capture details and textures in images, resulting in more realistic visual effects. This technology brings new possibilities and innovations to the field of image generation.

Currently, Stable diffusion has been successfully deployed on the TPU platform, with high computational efficiency and inference speed when generating low resolution images. However, with the development of display technology, in many cases, people often hope to obtain 2K, 4K, or even higher resolution images. The efficiency of directly using the Stable Diffusion model to generate high-resolution images is relatively low. A common method is to generate a low resolution image using the SD model, and then improve the resolution of the generated image through super-resolution reconstruction model, thereby improving the quality and details of the generated image. Therefore, deploying efficient super-resolution models on the TPU platform is of great assistance in generating high-resolution images using Stable Diffusion technology.

In order to improve the running speed of the stable diffusion on the edge computing box, reduce the resource consumption and memory occupation, this question proposes a scheme to use the super resolution reconstruction model as the upper sampler of the stable diffusion.

This topic is one of the competition topics. To view all topics, you can click to visit the official website of the 2023 CCF Big Data and Computing Intelligence Competition (CCF BDCI).

• Challenge tasks

This competition is dedicated to the implementation of super-resolution reconstruction models on the TPU platform. Participants choose to deploy the super-resolution model on the computable TPU chip. Participants need to balance accuracy and inference speed in the process of model deployment. The details are as follows:

1. Research on super-resolution reconstruction models:

Participants need to study and select suitable super-resolution reconstruction models to improve image clarity and level of detail. We can consider using the latest super-resolution reconstruction algorithms, such as convolutional neural network based methods (such as SRCNN, ESPCN, etc.).

2. TPU deployment:

Participants need to deploy the selected super-resolution reconstruction model to the TPU platform, including converting the model to a format acceptable to TPU, writing corresponding inference code, and performing model loading and inference operations on TPU.

3. Performance optimization:

Participants need to optimize the performance of the super-resolution reconstruction model to reduce running speed and ensure that the image reconstruction quality has lower NIQE indicators. Methods such as model compression, quantization, and graph optimization can be considered to improve inference speed.

The competition adopts a "two-level competition system" of preliminary and final rounds, and the specific schedule is as follows:

September 25, 2023- November 27, preliminary stage

1. On September 25, 2023 (12:00), the competition questions will be released. Players can log in to the competition official website to register;
2. On October 14, 2023, online evaluation for the preliminary round will be launched. Players can submit result files online to the competition platform, with a maximum of 3 submissions per team per day. The evaluation system will automatically evaluate the scores and update them to the leaderboard synchronously. The ranking will record the highest score of the contestant, and the relevant team must save the source code of the highest score work for review;
3. 2023/11/20 (12:00), deadline for registration and team formation;
4. 2023/11/25 (24:00), deadline for submission of A-list works in the preliminary round; After the preliminary A-list evaluation is completed, all contestants can participate in the preliminary B-list evaluation;
5. On November 26, 2023, the B-list test set will be announced. All participants in this competition must download the B-list test set on the "Question Data" page of this competition before 24:00 on November 26;
6. On November 27, 2023 (00:00-24:00), entries from the preliminary B-list will be submitted. Participants can submit their entries 3 times on the day of the B-list, but only the last submission of each team will be evaluated. The final qualification will be based on the online final score of the B-list (as displayed on the B-list leaderboard). If the team does not submit the B-list, they will not be able to advance to subsequent competitions. The TOP 5 team on the B-list will be qualified for the final after review and reproduction.

2023/11/28-12/17, Final stage

1. 2023/11/28-12/08, submission of reproduction and final materials. The team that plans to be nominated for the finals shall submit reproduction materials, final defense review materials, etc. as required;
2. December 15, 2023, final review (offline);
3. 2023/12/16-12/17, Final Carnival Series Activities (please follow the follow-up notice for details)

Remarks:

Teaming and submission of works should be done on the PC side;

The above schedule is calculated based on Beijing time, and the schedule may be adjusted according to the actual situation.

• Team Registration

1. Eligibility: Open to global teams - students, researchers, and professionals from various fields.
2. Registration: Multiple topic participation allowed, one team per topic. Identity verification required.
3. Team Formation: Teams limited to 1-5 members, unchangeable after formation. Submissions under team identity, with limits per team.
4. Responsibilities: Team leaders coordinate tasks and communicate with organizers.
5. Rules: Participants creating topics can't join them. Organizers and technical support can't participate or guide teams.

• Work Requirements

1. Originality: Entries must be original, abide by Chinese laws, and respect third-party rights. Violations lead to disqualification.
2. Intellectual Property: Rights are shared among sponsors, participants, and the platform. Organizers have priority for investment and promotional use.
3. Data Usage: Authorized data use for training only, not for commercial purposes. Proper data source acknowledgment required for research.
4. Compliance: Entries must comply. Violations like cheating, unauthorized data use, false submissions lead to disqualification and rearrangement of winners.

• Evaluation Rules

1. Evaluation: Eligible team entries are reviewed. The committee isn't responsible for computer or network issues affecting submissions.
2. Fair Play: Teams mustn't exploit loopholes or cheat. Violations lead to score cancellation.
3. Focus: Experts assess entries for maturity, innovation, practicality, societal impact, and commercial value.
4. Validation: Participants cooperate for entry validation. They ensure accuracy before submission as changes aren't allowed.
5. Result Confirmation: Expert decisions are final without further feedback.

• Notifications and Awards

1. Player Notifications: The committee contacts teams for event participation. No response within 3 days forfeits the opportunity, allowing other teams to join.
2. Prize Taxation: Winners cover taxes on prize money (in Chinese Yuan). The committee assists in tax payment.
3. Prize Distribution: The team leader receives the prize money and handles distribution among team members, with no responsibility held by the organizers.

• Organizer's Rights

1. The committee can refuse entry or submissions.
2. It can address violations and adjust awards for fairness.
3. Updates or rule changes may affect rankings.
4. The committee can modify event timelines.
5. They reserve the right to adjust team numbers, prizes, and deny awards.
6. Rules are subject to committee changes; final event interpretation lies with the organizers.