Face key point detection refers to accurately locating face contours and key points of facial features
under different expressions, genders, ages, postures, and ambient lighting conditions.
It is a basic function of the Driver Monitoring System (DMS).
It is also a key technology in the field of active safety that Streamax Technology focuses on.
ICME 2021 Reward
Recently, the Third 106 Face Key Point Positioning Competition that organized by ICME 2021 (IEEE International Conference on Multimedia and Expo) and JD Technology has officially ended. As the first face key point positioning with masks competition and digs for more efficient methods that can detect face key points with masks in the industry, its evaluation is more difficult compared with previous competitions. Due to the test samples cover a variety of complex scenarios such as large poses, extreme facial expressions, and mask occlusion, and also the fewer training samples, it brings huge challenges to the generalization and accuracy of the algorithm. While competing for positioning accuracy, it also pursues the practicality of the algorithm and has strict requirements on the complexity of the model. Therefore, the competition is of great academic and practical value. The Large Model Verification
There are 83 world-renowned universities and companies, such as Meituan, Tencent, ByteDance, University of Chinese Academy of Sciences, etc., that competed for more than a month. After verification, submission, and testing, the Streamax Technology Artificial Intelligence Center won first place in the large model verification and third place in the small model test.
The Large Model Verification
The Small Model Test
Interview
Today we invited the engineer representatives of the award-winning Streamax Artificial Intelligence Center, to talk about the stories of the competition.
Q1: Please introduce the original intention of this competition.
HU Hong: Due to the impact of the COVID-19, this face key point competition is very different from the previous two. The test set includes real masks and virtual masks data. Recognition of key points of faces with masks is also a problem that our team is working on. The drivers wearing masks pose a huge challenge to DMS (Driver Monitoring System), thus, we want to explore the face occlusion key point positioning. Moreover, the test stage of this competition has very strict limits on computational efficiency and model size, which is consistent with our commitment to front-end implementation and the pursuit of efficiency and performance.
Besides, every pilgrim has a Jerusalem in his heart. Before I joined the company, I also participated in international competitions for human body posture estimation before (COCO2020, LIP2019, FashionAI, etc.). With the help and support of the team and the mentality of learning, I would like to cooperate with colleagues and friends to participate in the competition for once.
Q2: What do you think are the important factors that lead to the award?
HU Hong: I think that on the one hand, it comes from departmental support and technical accumulation. More ideas will be generated after the collision of thoughts between colleagues, and the verification time of tricks can be greatly shortened. Competing in a contest is similar to working on a project: meeting the competition system in a limited time and improving the performance of the model method as much as possible. Our DMS team has been working on face technology for many years and that experience provides us with ideas and experience in the competition.
On the other hand, the competition needs a good baseline. Trying new methods on it will have achievement faster. Therefore, we have continued to use and expand the codebase of previous competitions and many attempts have been made.
Q3: Through this competition, what do you think you or your team have gained?
JIANG Tao: We have always been a relatively low-key company. There were not many people in the industry who knew our company team at the time of the competition. Through this competition, we are able to increase our influence in the industry to some extent. I think that during the competition, we have also enriched our technical accumulation on face key point tasks and provided more solutions to the scenario of drivers with masks.
HU Hong: Through this competition, we have more understanding that AI applications still face many difficulties, especially in the integrated landing of the front end, and recognize our shortcomings. In the past, most of the competitions did not require computing power. However, the restrictions on the number of model parameters and calculations in this competition made many methods in the past invalid, which further illustrates that AI technology is a difficult process from incubation to landing. As a newcomer in the workplace who has graduated for less than a year, I still need to further shape myself and accumulate in the DMS industry.
Q4: From your perspective, what are the positive impacts that can bring to the DMS team when participating in this competition?
HU Hong: There are strict requirements for the calculation and size of the model in the testing phase. Our team has achieved better results with only 0.2M parameters and less than 100M calculations without using quantization technology. The realization of the low computational load is very in line with the application scenarios of our DMS task and paves the way for the landing. Meanwhile, the robust model is trained in the verification phase, which provides prior knowledge for the automated annotation tools used by our team. Combined with the generation method of virtual masks, it provides an auxiliary means for the subsequent labeling of the data of masks in actual scenes, especially for improving the performance of tasks related to occlusion of faces.
In addition, the baseline of the key points of the human face has been accumulated and improved, which provides development and implementation tools for subsequent tasks.
Streamax Technology has always adhered to continuous innovation and attaches importance to technical application. The technology used in this competition is application-oriented, with the goal of solving user pain points and helping users create value. The technology has entered the stage of large-scale deployment and reserve and will have a linkage effect with other AI functions shortly. This will further enhance the core competitiveness of the products of Streamax Technology and enhance user experience.
In the future, Streamax Technology will continue to promote the development and innovation of core technologies in the field of computer vision. Through the continuous expansion of the application of related technologies, a more comprehensive and intelligent system terminal product will be formed, which will provide richer solutions for various industries.