On June 22, the Institute of Electrical and Electronic Engineers (IEEE) International Conference on Computer Vision and Pattern Recognition (CVPR) 2022 announced the Best Student Paper Award among other awards. Researchers from Tongji University and Alibaba won the Best Student Paper Award. The first author of this paper is CHEN Hansheng, a 2020 Master's degree candidate from the Institute of intelligent vehicles, School of Automotive Studies, Tongji University, and Shanghai Autonomous Intelligent Unmanned System Science Center. His supervisor is Professor XIONG Lu. The corresponding authors of this paper are TIAN Wei, an associate supervisor and assistant professor of School of Automotive Studies, Tongji University, and Dr. WANG Pichao, Alibaba. This was the first time that the first author of the award-winning paper was from a Chinese University since CVPR established the Best Student Paper Award in 2001.
Accurate pose estimation of objects has always been a difficult problem in autonomous driving. The paper entitled "EPro-PnP: generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation" focused on the problem of 6-DOF pose estimation of objects in monocular images, and proposed a general probabilistic N-point Perspective (PnP) algorithm, EPro-PnP (end-to-end probabilistic PnP). This algorithm can be inserted into the depth network as a differentiable module, output the pose probability density distribution on the SE3 manifold, and realize end-to-end object pose learning by minimizing the KL divergence between the predicted pose and the target pose distribution. EPro-PnP combines the advantages of traditional geometric vision and cutting-edge deep learning. It improves the accuracy of pose estimation in the 6-DOF object pose estimation task of line MOD dataset and the 3D object detection task of nuScenes dataset. It has a wide application prospect in robots and autonomous driving.
Professor XIONG Lu said that Tongji students' winning this award demonstrated the successful interdisciplinary cooperation between vehicle engineering and intelligent science and technology of Tongji University, and marked the achievements of integration between university enterprise and research. In the future, the team will continue to promote interdisciplinary research, enhance the cultivation of talents integrating production and education, and strive to contribute to the self-reliance and self-improvement of high-level science and technology in China.
The number of contributions to this year's CVPR conference reached 8161. After more than three months of review, a total of 2064 papers were accepted by the conference, with an acceptance rate of 25.28%.
CVPR is a top-level conference in the field of computer vision and pattern recognition organized by IEEE. After the Covid-19 pandemic broke out, CVPR was the first conference that was held offline this year. A total of 9981 people registered to attend the conference, including 5641 on-site participants and 4340 online participants.