On International Conference on Computer Vision 2017（ICCV’17）which was held on 29th Oct. in Venice, Megvii Research won four titles in Joint Workshop of the COCO and Places Challenges. In a total of seven challenges, Megvii Research participated four of them, and won three first-place and one second-place, outperforming other talents, including teams from Microsoft, Facebook, Google, and CMU. Megvii Face++ has also become the first Chinese company who wins first-place in COCO Challenges.
COCO Challenges Ranking Result
MS COCO (Microsoft Common Objects in Context) is a Challenge in computer vision and image recognition, which receives most attention after ImageNet challenge suspended. It is the only challenge which is able to gather all the well-known companies and research institutions globally, such as Microsoft, Google, Facebook, and many other universities around the world. Compared to ImageNet, COCO emphasizes more on detecting objects in an image, rather than classifying the image as what ImageNet focused. This requires more understandings of image details for algorithms, thus COCO is considered the highest standards challenge in succession to ImageNet.
COCO Challenge started from 2015, and the previous two years’ first-place winners for Detection Challenge is MSRA and Google. This year, it comprises four challenges: Detection Challenge, Instances segmentation, Keypoint Challenge and Stuff Segmentation. Megvii Face++ participated in first three challenges and won two first-place and one second-place with honor.
Places Challenge 2017 is a new challenge held jointly with COCO Challenge at ICCV’17, it hosts another three new tracks which meant to complement the COCO Challenges. The three specific tracks in the Places Challenge are: Scene parsing, Instance Segmentation, and Semantic Boundary Detection. Megvii Face++ participated Instance Segmentation and took the first place, defeating our main competitor, Google.
For many industries, machine recognition toward human, stuffs and scenes are of utmost importance. This achievement in COCO and Places Challenges proves that Megvii Face++ is leading the technologies globally. Our advantages are in competitive technologies, Megvii Research and product centers will continue to promote technological transformation into product development, to carry out behavior recognition, scene segmentation, object detection and object segmentation technology in fin-tech, intelligent security, smart city, new retail, mobile phones and many other scenarios or industry applications that are yet to explore, in order to maximize the value of Face++ technologies.
Jian Sun, chief scientist and director of Megvii Research, said that it is the following three core competencies that contribute to such accomplishment in this competitive challenge:
1. Credits to our self-developed learning engine, MegBrain, which allows us to systematically train the algorithm at the fastest speed.
2. Megvii Research has a long and deep study of deep learning and computer vision algorithms. Although we are rarely seen at international data set competitions, but in fact the internal technical indexes have always been very high and competitive. We take this opportunity to present ourselves and share with you what Face++ are really capable of.
3. In addition to having abundant computing resources, Megvii Face++ boosts to have an environment that encourages constant high-speed innovation and a culture that is always in pursuit of the ultimate.
At last, Jian said, “congratulations to our team at Venice, we are so proud of you.”
