After two months of fierce competition, Video++ (极链科技), an artificial intelligence (AI) company on China’s video AI scene, won the first place in the Google Landmark Recognition 2019. This is the second time Video++ won one of the world’s most prestigious AI competitions, another milestone for the Chinese AI team in international competition.
As the winner of this competition, Video++ will once again be invited to the 2019 Conference on Computer Vision and Pattern Recognition (CVPR) held in the United States from June 15 to June 21and the company will be participating in the world’s top seminar on computer vision.
Last year, Google released Google-Landmarks, so far the largest landmark data sets worldwide. In order to boost development in instance-level recognition and image retrieval, Google also launched two competitions, namely, Landmark Recognition 2018 and Landmark Retrieval 2018, with more than 500 researchers in machine learning taking part.
The contestants of Google Landmark Recognition 2019 signed up and submitted on the Kaggle platform, the world's most authoritative data science competition platform and has millions of data scientist users worldwide. The rules of entry require that the contestant recognize all images containing the query landmarks in the database after being given a certain query image. With million-scale data volume and image content containing too much irrelevant information, the competition is extremely difficult.
In addition, Google unveiled its Landmarks v2 this year. The data set contains 4.13 million images, more than three times those last year. The number of different landmarks has reached 200,000. Based on the new data set, the Google Landmark Recognition 2019 attracted 281 teams and individuals experienced in computer vision to further detect and train their models with more complex landmarks.
Surpassing many well-known AI teams in the world, Video++ ranked first with the highest score of 0.37606, while the second and third place scored 0.35988 and 0.35541 respectively.
According to the winning strategy shared by contestants from Video++, their algorithm is composed of two models and three steps in the application. The detailed algorithm will be uploaded to arXiv in the form of an essay.
The AI team of Video++ also stated that the algorithms they used in this competition will be applied in its scene recognition and video recognition products.
Currently, Video++ has grown into a unicorn with six rounds of funding in the past four years, led by investors including Alibaba Group, YF Capital, UBTECH, Megvii, Sirius Capital, Winshare Capital and so on. It has altered the traditional form of internet video ad placement through AI and made big progress on the efficiency of video ads. Facial AI and Video OS are two core products in computer vision application.