For the first time in more than 20 years! CVPR best student thesis awarded to Chinese college students!

The most influential conference in the field of global artificial intelligence CVPR Recently announced , take 2022 year “ Best student paper ” The award was given to chenhansheng, a second year graduate student of Tongji University , This is the first time in more than 20 years that Chinese college students have won this honor . This paper is the result of Chen Hansheng's internship at Alibaba Dharma academy , It combines traditional geometric reasoning and deep learning , Put forward the calculation 3D Object pose （ Position and orientation ） New method of , Can calculate from a single picture 3D The position and orientation of an object in the real world , It is expected to become automatic driving 、 The theoretical impetus for the development of robotics and other technologies .

CVPR（ International Conference on computer vision and pattern recognition ） It is the top academic conference in the field of artificial intelligence , This year, we received 8161 Papers submitted , Final acceptance 2064 piece , Only the best thesis and the best student thesis will be selected 1 piece .CVPR The best thesis and the best student thesis are often regarded as AI Technology weathervane , The birth of ResNet And so on .

CVPR As early as 2001 The best student thesis award was established in , This year, it will be distributed to Chinese college students for the first time . Award winning papers 《EPro-PnP: Generalized End-to-End Probabilistic Perspective-n-Points for Monocular Object Pose Estimation 》 From the automobile College of Tongji University and Alibaba Dharma Academy , The authors are chenhansheng 、 Wangpichao 、 Wang Fan 、 Tian Wei 、 Xiong Lu 、 Li Hao .

award certificate

The thesis studies 3D Classic problems in the field of vision , Solve through a single image 3D The position and posture of objects in the real world （ Position and orientation ）. This technology has a broad application prospect , It's self driving 、 Basic technology of robotics and other industries . For example, in automatic driving , Only calculate the position and posture of the surrounding vehicles first , Judge whether the other party wants to speed up 、 Braking or changing lanes , Only own vehicles can perform corresponding operations , Inaccurate or too slow calculation may lead to accidents .

3D Schematic diagram for calculating the pose of an object

Locate through a single image 3D Objects are very challenging . There are generally two types of solutions ：

• One is based on geometric reasoning , for example PnP Algorithm , Good explainability 、 Strong generalization ability , But you need to know the size and shape of the object in advance , It has great limitations ;
• The other is the deep learning method , Predictable 3D Position coordinates and orientation angle of the object , But it is easy to over fit on small data sets .

The new method proposed by the award-winning paper EPro-PnP, Creatively introduced the probability distribution , The two methods of geometric reasoning and deep learning are seamlessly connected , Formed an end-to-end easy-to-use model , Can quickly estimate 3D The pose of an object . Experimental proof , The new model is universal 、 Accurate positioning , There is no need to know the geometry of the object in advance ; More importantly, it is very concise , More efficient , It has good interpretability , It is expected to be used for automatic driving 、 robot 、 Unmanned aerial vehicle (uav) 、AR And many other scenes that need to estimate the pose of objects through vision .

EPro-PnP Method diagram

according to the understanding of , Chen Hansheng, the first author of the paper, studied in the school of automotive engineering, Tongji University , At present, the second grade graduate student is studying , The tutor is Professor Xiong Lu , The associate tutor is assistant professor Tianwei , The tutor of Dharma hall is Dr. wangpichao . His research interests are in computer vision 3D Object pose estimation , The first research has been in CVPR publish one’s thesis .2021 After working as a research intern at the Ali Dharma Institute in , Inspiration comes from daily research and discussion in Dharma hall , Try to construct a unified theoretical framework .“ I thought this paper might be quite popular , Because it is too mathematical , I didn't expect to win the prize .” Chenhansheng said .

Research Intern of Dharma Academy

Chenhansheng, a graduate student of Tongji University, is the first corresponding author of the thesis 、 Dr. wangpichao, an algorithm expert of Dharma Institute, said , What is special about this paper is that it is solved mathematically AI Underlying problems , It has brought a breakthrough in basic theory , Yes 3D The research and application in the field of vision will have a positive impact .“ The core innovation is to change the perspective of mathematics , Introduce probability density , Change a non differentiable function into a differentiable function , So we can train the deep network through back propagation , Finally, stable convergence is achieved , promote 3D Positioning accuracy of objects .”

The research intern is from Alibaba 2017 The scientific research project founded in , Up to now , For the world 200 Many top universities 1000 Many students have provided scientific research positions .

【 Thesis link 】https://arxiv.org/abs/2203.13254

For the first time in more than 20 years ！CVPR The best student thesis is awarded to Chinese college students ！