CVPR 2022 remembers Sun Jian! Tongji and Ali won the best student thesis award, and hekaiming was shortlisted

---

New Zhiyuan Report   

edit ： Editorial Office

【 Introduction to new wisdom 】2 I haven't seen you in years ,CVPR 2022 Finally came offline ,6000 Ginseng will be popular . The best paper goes to Zurich Institute of technology , Tongji 、 Ali won the best student thesis . Professor lifeifei of Stanford University won 「 Huangxutao Memorial Award 」. Besides , A video was also broadcast at the conference , In memory of Dr. Sun Jian .

6 month 21 Japan ,CVPR 2022 Officially open the curtain .

After two years of virtual meetings affected by the epidemic , The conference finally restored the way of offline attendance .

therefore , The number of offline registered participants suddenly reached 5641 people , Than 2017 More than a year .

From the pictures sent from the scene , It is really a sea of people .

CVPR 2022 Received it from 23389 The authors are 8161 Contributions , comparison 2021 Year of 7093 Article growth 15%.

among , The authors from Chinese Mainland contributed the most , Accounted for 44.59%; The United States uses 20.65% In the second place .

After the Organizing Committee 3 More than months of detailed review work , The conference finally accepted 2064 Papers , The reception rate is about 25.3%, Higher than last year's 23.6%.

among , Yes 342 This paper was selected as Oral The paper ,1721 Selected as Poster The paper .

Speaking of this , I have to mention the quality of manuscript review . Recent years , Whenever academic conferences are released , Various roast will come one after another like a tide .

such as , Some time ago, I was pushed to the forefront of the storm ICML 2022, Namely LeCun I submitted three articles and was rejected 3 The one in Chapter , After the general operation of the conference, Professor Ma Yi was so angry that he said he would not contribute any more .

Regarding this ,CVPR It emphatically explained its manuscript review mechanism and achievements .

On the experience distribution of reviewers , Current CVPR Proportion of student reviewers 31%, Teaching post / The proportion of reviewers of researchers is 69%.

among ,23% Of the reviewers participated in the review less than twice . Participating in the review is greater than 10 The proportion of senior reviewers for the second time is 19%.

It can be said that the perfect practice 「 Three combinations of the old, the middle-aged and the young 」.

On the task assignment of reviewers , Basically, the more experienced reviewers , The more papers you review , The heavier the task , Review experience is greater than 10 On average, the senior researchers of times have to review 5.7 piece , Student reviewers who have less than two times of experience in reviewing manuscripts , On average, it takes only 2 piece .

Reviewers of academic conferences are not paid , To thank them for their selfless dedication , Current CVPR It's been judged 100 Many outstanding reviewers , And awarded them the value 100 A gift certificate for us dollars .

（ There's applause ）

Besides , The General Assembly paid special attention to 「Rebuttal How useful 」 Made statistics .

data display , There are nearly 3000 Papers are in Rebuttal after , The highest score was reduced 0.5 branch , Yes 660 The number of articles decreased at most 1 branch .

Because the lower the score , The higher the evaluation , So on the whole ,Rebuttal It has played a positive role in the acceptance of the thesis , Of course, there are risks ,Rebuttal After that, the evaluation became worse 2000 Multiple articles .

Current CVPR The review process of the conference is as follows ：

Each paper will be written by 3 position AC Deal with it together , Including at least one senior AC And at least two experts in the field ,AC The assignment of is generated by automatic matching .

Each paper will be assigned at least 3 Three reviewers , At least two of them are main AC recommend , Lord AC Preside over the reviewer's meeting to discuss .

AC Hold an online meeting to discuss the reviewers' opinions in detail , And review the reports with each other , Check for errors , Give notice to the author .

in addition , The meeting is also for all AC Prepared training materials for reviewers .

CVPR 2022 Awards

Best Paper Award ：

「Learning to Solve Hard Minimal Problems」

Address of thesis ：https://openaccess.thecvf.com/content/CVPR2022/papers/Hruby_Learning_To_Solve_Hard_Minimal_Problems_CVPR_2022_paper.pdf

Institutions ： Zurich Federal Institute of Technology 、 University of Washington 、 Georgia institute of technology 、 Czech University of Technology

This paper presents a new method in RANSAC Methods for solving difficult geometric optimization problems within the framework . This kind of problem involves relaxing the original geometric optimization problem to a minimum problem with many false solutions . The method proposed in this paper avoids calculating a large number of false solutions . This paper proposes a learning strategy , Used to select the problem that can continue the original problem numerically , And the starting problem of the solution of interest - The solution to .

This passage RANSAC Solver to prove this method , The solver is used to calculate the relative positions of three calibrated cameras , Minimum relaxation through four points in each view . The average is in 70 Solve a problem in microseconds .

Honorary nomination for best paper ：

「Dual-Shutter Optical Vibration Sensing」

Address of thesis ：

https://openaccess.thecvf.com/content/CVPR2022/papers/Sheinin_Dual-Shutter_Optical_Vibration_Sensing_CVPR_2022_paper.pdf

Institutions ： Carnegie Mellon University

Visual vibration measurement is a very useful tool , It can be used to capture audio remotely 、 Physical properties of materials 、 Human heart rate, etc . Although visually observable vibrations can be captured directly by high-speed cameras , But through the displacement imaging of speckle pattern generated by laser beam irradiating the vibrating surface , It can amplify the vibration of small and imperceptible objects optically .

This paper presents a new method , Used to sense high speed （ the height is 63kHz） The vibration of , It can sense multiple scene sources at the same time , The rated operating frequency of the sensor used is only 130Hz.

This method relies on two cameras equipped with rolling and global shutter sensors to capture the scene at the same time . The rolling shutter camera captures distorted speckle images , Encode the vibration of high-speed objects . The global shutter camera captures an undistorted image of the speckle pattern as a reference . By capturing the sound source （ Such as loudspeakers 、 Voices and instruments ） Vibration caused by , Analyze the vibration mode of tuning fork , This new method is demonstrated .

Best student thesis Award ：

「EPro-PnP: Generalized End-to-End Probabilistic Perspective-n-Points for Monocular Object Pose Estimation」

Address of thesis ：

https://openaccess.thecvf.com/content/CVPR2022/papers/Chen_EPro-PnP_Generalized_End-to-End_Probabilistic_Perspective-N-Points_for_Monocular_Object_Pose_Estimation_CVPR_2022_paper.pdf

Institutions ： Tongji University 、 Alibaba Group

adopt Perspective-n-Points（PnP） From a single RGB Positioning in the image 3D Object is a long-standing problem in computer vision . Driven by end-to-end deep learning , Recent research suggests that PnP Interpreted as a differentiable layer , The gradient of the pose of the object can be propagated back , To learn some 2D Images and 3D Correspondence of spatial points . However , If you start from scratch and learn the whole unlimited 2D-3D Point set , There is no convergence in the current method , Because the definite posture itself is non differentiable .

This paper proposes Pro-PnP, A probability for general end-to-end pose estimation PnP layer , It can output SE(3) The distribution of postures on manifolds , In fact, it will be classified into Softmax Bring into the continuous field .2D-3D The coordinates and the corresponding weights are treated as intermediate variables , By minimizing the relationship between prediction and target pose distribution KL Spread out to learn . Its basic principle unifies the existing methods , It's similar to the attention mechanism .

EPro-PnP The effect is obviously better than the existing benchmark , stay LineMOD 6DoF Pose estimation and nuScenes 3D object detection benchmark , Reduced based on PnP The gap between the method of and the best performance for a specific task .

Honorary nomination for best student thesis ：

「Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields」

Address of thesis ：https://arxiv.org/pdf/2112.03907.pdf

Institutions ： Harvard University, 、 Google research

Neural radiation fields （NeRF） Is a popular view synthesis technology , It represents the scene as a continuous volume function , The volume density of each position and the parameters of the emitted radiation varying with the line of sight are provided by the multilayer perceptron . Although based on NeRF Our technology is good at representing fine geometric structures with a smoothly changing line of sight dependent appearance , However, it is often unable to accurately capture and reproduce the appearance of glossy surfaces .

In this paper, by introducing Ref-NeRF To solve this problem .Ref-NeRF Replaced by the expression of reflected radiation NeRF Parameterization of line of sight dependent outgoing radiation , A series of spatially varying scene attributes are used to construct this function .

It turns out that , Combined with the regularizer of normal vector , The new model greatly improves the authenticity and accuracy of specular reflection . Besides , The model in this paper can explain the internal representation of the emittance , Very useful for scene editing .

33 A shortlist of the best papers

The first one in the shortlist is hekaiming's MAE, But it's a pity. , No, I didn't get a rating at last .

As early as 2021 year MAE Time of birth , Hekaiming is highly expected by everyone .

National Physics 、 First prize in chemistry competition , Send Tsinghua University , Full marks in the college entrance examination , Microsoft Asia Research Institute internship ,CVPR The first Chinese winner of the best Thesis Award ...

so to speak , Any one of these proud resumes can kill most of their peers .

Although he Kaiming did not win this CVPR Best paper , But he will make greater achievements in the future .

Longuet-Higgins prize

Longuet-Higgins Award by IEEE Of PAMI（ Pattern analysis and machine intelligence ,pattern analysis and machine intelligence） Technical Committee , Every year CVPR Post on the Forum , In recognition of CV（ Computer vision ） Having a major impact on research in the field CVPR The paper .

The prize is awarded to theoretical chemists and cognitive scientists H. Christopher Longuet-Higgins To name , The winner is TCPAMI Another committee appointed by the awards committee will decide .

This year, ,Longuet-Higgins Award by Andreas Geiger And other three authors 「 Are we ready for autopilot ？——KITTI Visual reference assembly 」 get . Of the three authors ,Geiger and Lenz Karlsruhr Institute of Technology , and Urtasun From Toyota University of technology Chicago .

In this paper , The three researchers used the self-developed autopilot platform as a three-dimensional 、 Optical flow 、 Visual range / SLAM and 3D A new benchmark has been developed for target detection and other tasks .

among , The researchers' recording platform is equipped with four high-resolution video cameras 、Velodyne Laser scanner , as well as SOTA Level localization system .

Young Researcher Award

The young researcher award is also awarded by PAMI Issued by the technical committee , The selection process is conducted by PAMI-TC The awards committee manages . The winner will also be in CVPR It was announced at the meeting that .

There are two winners this year , Namely Bharath Hariharan and Olga Russakovsky.

Hariharan He is an assistant professor of computer science at Cornell University , Mainly engaged in CV and ML Work on , Especially those who ignore 「 big data 」 Important issues of labels .

He likes to solve the problem of integrating the progress of machine learning with computer vision 、 The combination of geometry and domain specific knowledge .

at present ,Hariharan Our team is building a system that can learn tens of thousands of visual concepts with little or no supervision , And produce detailed output , For example, precise 3D Shapes and so on .

Russakovsky He is an assistant professor of computer science at Princeton University , The research direction also focuses on CV field .

Huang xutao （Thomas S. Huang ） A Memorial Award

Huangxutao Memorial Award was awarded on 2020 It was set up in , from 2021 Year begins , Every year CVPR One time at the general meeting , In recognition of the research 、 Teaching and guidance and for CV Researchers recognized as exemplary in community service .

The award was established in memory of the late professor huangxutao , He is a teacher in CV And image processing . meanwhile , He is also an example of contribution to the growth of several generations .

This award is awarded annually to at least 7 Researchers in , It's better to be in the middle of your career （ A doctor's degree cannot exceed 25 year ）.

Candidates for the award consider CV All researchers in the field . The final reward includes 3,000 Dollars in cash and a medal .

The winner of this huangxutao Memorial Award is lifeifei, the first Sequoia professor in the computer science department of Stanford University .

2013~2018 year , She served as director of the Artificial Intelligence Laboratory at Stanford University .2017 year 1 Month to 2018 year 9 During the month , She also served as vice president of Google , Served as the chief scientist of artificial intelligence and machine learning in Google cloud .

Dr. lifeifei devotes himself to the study of heuristics AI、 machine learning 、 Deep learning 、 Computer vision , as well as AI+ Health care .

Besides , She has also studied cognitive and computational neuroscience .

She was Nature、PNAS、CVPR、ICCV、ECCV、ICRA And so on 200 Many papers .

She is still ImageNet and ImageNet Challenge The inventor of the , This is a large-scale data set and benchmark work , It has made great contributions to the latest development of deep learning and artificial intelligence .

Besides her technical achievements , She is still STEM Advocate of ,AI A national spokesman for diversity . She is committed to improving the inclusiveness and diversity of AI education .

In memory of Dr. Sun Jian

2022 year 6 month 14 Early morning , Sun Jian, the chief scientist of Kuangshi, passed away .

CVPR 2022 Conference site , Before publishing the best paper , In memory of Dr. SunJian .

A video on the scene led us to recall Dr. Sun Jian's scientific research and growth path .

‍

‍

‍

‍

Last , Daniel of artificial intelligence community 、 And Sun Jian's relatives and friends have written a memorial letter to Dr. Sun Jian .

Reference material ：

https://twitter.com/CVPR