Engineer 9x Viettel won the international AI contest thanks to simple thinking

This year’s AI City Challenge, with the participation of many major competitors in the world such as Baidu, Tencent, Oppo, Alibaba…, is a contest to promote the application of artificial intelligence (AI) in improving the operational efficiency of urban environment. Each year, the competition will present specific problems, the teams will research the solution and be ranked based on the results that the solution brings. This is the first time VTCC has participated in this contest.

Find solutions from simple things

In 2022, in the category of Naturalistic driving analysis, the competition presented 90 videos containing 18 distracting behaviors of drivers and asked the teams to find a solution to accurately identify those behaviors at the right time. .

Although they have participated in a number of international computer science competitions such as Kaggle, Quan and Tung are still quite strange when receiving this topic from AI City Challenge. Before that, your group mainly focused on behavior recognition on photos, never deployed on video. The time you received the information about participating in the contest was also quite late: only 1 month before the deadline, equivalent to 1/3 of the time the contest allowed.

Engineer Vu Minh Quan (left), Engineer Tran Manh Tung (right)

A completely new problem solved in a very short time, Quan and Tung were forced to find an optimal direction.

“Preliminary assessment of the requirements for the title, we decided to choose the simplest approach: divide the video into small segments, identify the behavior on each segment, and then combine the results. For small-scale and less diverse data sets like the contest data set, this is an effective approach that avoids rote learning of the AI ​​model, leading to inaccurate results when applied in different situations. larger data set,” Quan shared.

The phase of fine-tuning the solution that took several months to deploy is now encapsulated in 7 days. Errors kept appearing, the result points were almost lost in the rankings.

“Before the exam, we set a goal to be in the top 6. But until 1-2 days before the registration gate closed, we still couldn’t reach the top 10,” Quan said.

Persevering with the chosen method, Quan and Tung received perfect results, the score of the test jumped to the top 1 right after the preliminary round, the accuracy of behavior recognition was the highest among the teams even when applications with larger data sets. The team’s solution was also selected as one of 28 solutions presented at the 6th AI City Challenge Conference, within the framework of the CVPR Computer Vision and Pattern Recognition Conference held in the United States at the end of the year. June 2022.

“If you’ve done it, be determined to do it”

Although they are “hard friends” at work, sharing many similar views and perspectives, Quan and Tung have completely opposite attitudes at work. Tung said: “I never feel confused or worried about the results of the contest, even when the time is short or the math is difficult. If you’ve done it, just be determined to do it.”

As for Quan: “Or worry a lot. In order not to be discouraged or too pressured at work, I usually set a big goal, then break it down into smaller goals. Then temporarily forget about that big goal and focus on each small goal. That way, I will always feel motivated to continue.”

That’s the difference, but Tung and Quan all have one thing in common: persistence with work. 2 days before the registration deadline, the score is still very low, the two of you take turns checking and correcting the error. There are errors that take 1 day, all from input to output, to be found. To optimize time, with any error, you also thoroughly apply the simplest approach to handling. At the same time, research additional solutions to push the score higher. A few hours before the deadline, the two of you were still correcting mistakes, tweaking the model, and only stopped when the submissions were completely exhausted.

The last week of the sprint, the two engineers worked overtime every day in the evening, taking advantage of every minute to improve the quality of the solution. Recalling the days when he often left the company at 8-9 pm, Quan smiled, “Fortunately, I don’t have a lover”.

Motivation to stick from the element “public officials”

The factor that made Quan stick with VTCC was his colleagues in the Center. “Everyone here is very young and fun. I can share, and I also find a good partner at work.”

As for Tung, “What I remember most is that during the preparation of the test, Mr. Hoang Ngoc Duong, Deputy Director of the Center constantly encouraged the group. There are days at 11 am, he even texted “Try to be uncle”. I feel accompanied.”

VTCC . engineer team

At VTCC, young people like Tung and Quan are given all conditions to improve their own capacity. Participating in competitions is part of the job, and you can spend all your time researching and pursuing new techniques. The center even sent more personnel with real combat experience, skills and knowledge to support.

That is also the reason why Mr. Bui Khac Hoai Nam, Principal Engineer of Technology Solutions, VTCC, was arranged to guide and accompany Quan and Tung during the exam, even though working in a completely different field. . With experience in participating in many international arenas including AI City Challenge, Mr. Nam has directly given you many useful tips and tricks.

The orientation of becoming a leading unit in AI research and application, improving the quality of human resources in the field of AI technology continues to be the top priority of VTCC in the coming time. The Center will continue to create conditions for young employees to increase their friction and learn on an international scale through competitions, awards, training opportunities, etc. At the same time, it will give you the opportunity to research Research and apply new technologies to Viettel’s AI products, contributing to the future of digital life.

Xuan Thach


Leave a Reply

Your email address will not be published. Required fields are marked *