From June 19 to 24, CVPR 2022 (computer vision and pattern recognition), the world's top international conference in the field of artificial intelligence computer vision, was held in New Orleans, USA, and the online conference was held simultaneously. A total of 71 papers from Shangtang technology and joint laboratory were selected for this CVPR, reaching a new high. Nearly a quarter of them were recruited as oral reports, covering many frontier research fields and directions that have attracted much attention, such as 3D vision and autonomous driving, and continue to consolidate the leading momentum in the field of computer vision research in the world.
Since its establishment, Shangtang and United laboratories have published more than 700 papers at various academic summits and won more than 70 World Championships in various competitions. At the same time, Shangtang has not taken the number of papers as the standard to measure the technological development of the company. Dr. Wang Xiaogang, co-founder of Shangtang technology and President of the Research Institute, said: "we hope to encourage and guide young researchers to do high-level and high-quality research from the perspective of solving practical problems in the industry by providing a good environment for scientific researchers to conduct efficient research."
At this CVPR, Shangtang technology also participated in a number of academic competitions and also made remarkable achievements. For example, Shangtang technology, together with the Institute of automation of the Chinese Academy of Sciences and the Shanghai Artificial Intelligence Laboratory, participated in the embodied AI 2022 (2022 embodied intelligence challenge) and won the championship at the RXR habitat track. As an authoritative competition in the field of global embodied AI research, the competition requires natural language control to solve the navigation problem of indoor robots. Shang Tang's method achieved more than 90% improvement in the effect, the navigation accuracy increased from 24.08% to 45.82%, and the navigation fidelity increased from 37.39% to 55.43%. At the same time, in the Clic (challenge on learned image compression) competition held to promote the visual coding technology based on deep learning, the scheme provided by Shangtang technology team successfully won the champion of the image coding track, which not only achieved the best subjective evaluation score in all three test code points, but also had the fastest decoding speed in all deep learning schemes.
Promote technology enabled industries and lead industry breakthroughs with innovation
Shangtang technology has always encouraged research teams to pay attention to industrial needs and pain points, and combine research work with actual business scenarios. In recent years, relying on the construction of AI infrastructure such as sensecore Shangtang AI devices, Shangtang has stronger support in frontier research fields, further promoting the deepening of collaboration with industry, and leading the development of the industry with AI technology innovation.
For example, in the paper "bailando: 3D dance generation via actor critical GPT with choreographic memory", researchers proposed a new music to dance framework bailando, which can drive 3D characters to dance with music, and can not only ensure the standard and beauty of movement, but also maintain the consistency with different musical rhythms in time. At present, in the context of the improvement of AI, cloud computing and other technical capabilities, the application scope of digital people is becoming richer and richer, and they are gradually integrated into our lives in social networking, games, live broadcasting, virtual idols and other fields. This research undoubtedly provides a potential direction for the future digital human industry to shape more intelligent and personalized characters to meet diversified needs.
Pttr diagram of point cloud tracking framework
In recent years, with the development of autopilot and lidar technology, target tracking based on point cloud has also received more attention. In view of the unique challenges of point cloud data and the defects of existing algorithms, in the paper pttr: relational 3D point cloud object tracking with transformer, Shangtang research team proposed a novel point cloud tracking framework pttr, which significantly improved the accuracy of target tracking on multiple data sets and laid a foundation for the safe operation of autonomous driving.
Shangtang technology also jointly held a robust machine learning competition for complex scenes - robust models towards open world classification with the team of Professor liuxianglong from Beijing University of Aeronautics and Astronautics. The competition aims to promote the research on safe and reliable AI models, encourage the creation of safer and more reliable AI, and support the more sustainable development of artificial intelligence technology. The competition attracted 286 teams and 416 contestants. On June 19, the winners of the competition were officially announced on cvpr2022 art of robotics workshop.
Strengthen infrastructure and ecological construction and help generate achievements
The outstanding achievements of Shangtang technology in academic research and technological innovation cannot be separated from the strong computing power foundation and leading algorithm ability of leading software and hardware infrastructure integration, as well as the long-term accumulation of Shangtang in the construction of academic ecology and open source ecology. Shangtang provides important basic support for technology research and development and implementation by building and continuously improving the infrastructure with sensecore Shangtang AI large device as the core. Researchers can efficiently conduct scientific research, quickly experiment and verify new ideas, accelerate innovation and iteration, promote the production of high-level papers, and solve problems in the industrial implementation.
Shang Tang attaches great importance to the construction of academic ecology. Since 2017, Shangtang technology has successively established joint research institutes or laboratories with Shanghai Jiaotong University, Nanyang University of technology and Zhejiang University, established a special plan for the deep integration of industry, University and research with Tsinghua University, and promoted the establishment of a global academic alliance of artificial intelligence universities, so as to promote the production of various academic achievements and international academic exchanges and cooperation through close contact with the academic community. On June 11 this year, Shangtang technology and the global AI academic alliance of colleges and universities successfully held the "endless research: Shangtang paper sharing meeting", which brought together researchers and guests from Shangtang technology and Chinese University of Hong Kong, Zhejiang University, Nanyang Technological University, Peking University and other universities to interpret CVPR 2022 oral papers in the fields of 3D vision, posture estimation, bottom vision, representation learning, scene understanding online, Share valuable academic experience.
In addition, Shangtang continued to consolidate the construction of open source ecosystem. Openmmlab, an open source project based on visual algorithms, currently has more than 50000 stars in GitHub, and has successfully opened thousands of models to researchers and the industry. Opendilab, based on decision intelligence, was released at the waic conference last year and is open source to academia and industry. In the direction of big model, Shangtang, together with Shanghai Artificial Intelligence Laboratory and colleges and universities, jointly released the universal vision technology system scholar Internet, and open source opengvlab to help the basic research and ecological construction of universal artificial intelligence. Openmmlab also held a seminar on the theme "openmmlab: basic platform for computer vision research and production" during CVPR, and invited academic celebrities to participate in sharing and discussion to jointly build an open source ecosystem.
With the construction and improvement of infrastructure and the cultivation of academic and open source ecology, the foundation of artificial intelligence technology research will be more stable and broader. Shangtang will take this as a support to continue to lead the innovation of artificial intelligence technology, strengthen the deepening of the implementation of AI industry, accelerate the large-scale application, and promote the continuous breakthrough of artificial intelligence technology and industrial development.