Tongji and Ali won the CVPR best student thesis, lifeifei won the Huang xutao award, and nearly 6000 people attended the offline conference

Almost Human reports

Machine center editorial department

just ,CVPR 2022 The best paper was officially published 、 Best student thesis, etc . come from ETH Zurich、 University of Washington 、 Georgia institute of technology 、 Researchers from several institutions of the Czech University of technology jointly won the best paper award , Researchers from Alibaba and Tongji University won the best student thesis award . Besides , Professor lifeifei of Stanford University won the award of Thomas S. Huang （ Huang xutao ） A Memorial Award .

‍

2022 year ,‍‍CVPR The number of contributions to the conference reached 8161 Share , comparison 2021 Annual 7093 Submissions increased 15%. among 44.59% The author is from China , In second place is the United States , occupy 20.65%.

‍

Picture source ： Twitter users @Jared Heinly.

After more than three months of review , In the end, the conference has 2064 Papers were received , The receiving rate is 25.28%, Higher than last year's 23.6%. Among them is 342 Were selected as Oral The paper ,1721 Were selected as Poster The paper . Besides , This year's review The number of 25804,rebuttal The number of 5884.

‍

Picture source ： Twitter users @raven.

among , A large number of papers have been accepted in China . According to the heart of the machine , Shangtang technology and joint laboratory Co., Ltd 71 Papers were accepted by the conference , Nearly a quarter of them are employed Oral.

Since the epidemic of Xinguan , This year, CVPR It is also the first time to hold offline . According to the organizer's statistics , As of the opening of the conference , share 9981 People registered to attend the meeting , among 5641 People attend the meeting on site ,4340 People attend the meeting online .

This year's Congress specially commemorated Dr. SunJian ：

‍

At the beginning of the main meeting ,CVPR 2022 That is to say, all the awards of this conference have been announced . come from ETH Zurich、 University of Washington 、 Georgia institute of technology 、 Researchers from several institutions of the Czech University of technology jointly won the best paper award , Researchers from Alibaba and Tongji University won the best student thesis award . Besides , The awards for best thesis nomination and best student thesis nomination are respectively awarded by CMU And Harvard University 、 Google research received .

Best paper

• Institutions ：ETH Zurich、 University of Washington 、 Georgia institute of technology 、 Czech University of Technology

• Address of thesis ：https://arxiv.org/abs/2112.03424

Abstract of paper ： This study proposes a new method of RANSAC A method for solving difficult geometric optimization problems in framework . The minimization problem originates from the relaxation of the original geometric optimization problem （relax） For the smallest problem with many false solutions . The method proposed in this study avoids calculating a large number of false solutions . The researcher designed a learning strategy , Used to select the initial question - The solution is to continue to solve the original problem by numerical method . This research is done by creating a RANSAC Solver to demonstrate the proposed method , The solver uses... In each view 4 Points are calculated by minimum relaxation 3 The relative position and pose of two calibrated cameras . On average, , This method can be used in 70 μs Solve an original problem within . Besides , This research also makes a benchmark test and Research on the problem of calibrating the relative pose of the camera .

Best paper nomination

This year's best paper was nominated by CMU The study of 《Dual-Shutter Optical Vibration Sensing》 get .

• Institutions ：CMU

• Address of thesis ：https://openaccess.thecvf.com/content/CVPR2022/papers/Sheinin_Dual-Shutter_Optical_Vibration_Sensing_CVPR_2022_paper.pdf

Abstract of paper ： 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 .

In this paper , The researchers have proposed a new method in high speed （ the height is 63kHz） A new method to detect the vibration of multiple scene sources at the same time , The method uses a rated operating frequency of only 130Hz Sensor . Their method uses two cameras equipped with rolling and global shutter sensors to capture the scene at the same time , The rolling shutter camera captures a distorted speckle image encoding the vibration of a high-speed object , The global shutter camera captures an undistorted reference image of the speckle pattern , Thus, it is helpful to decode the source vibration . Last , By capturing the audio source （ Such as loudspeakers 、 Voices and instruments ） And analyze the vibration mode of tuning fork , Showing their approach .

The researchers used a novel method 「 notice 」（seeing） voice .

Best student paper

This year, the winner of the best student thesis award is 《EPro-PnP: Generalized End-to-End Probabilistic Perspective-n-Points for Monocular Object Pose Estimation》, The author is from Alibaba 、 Tongji University .

• Institutions ： Alibaba 、 Tongji University

• 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

Abstract of paper ： Use perspective points （PnP） Cardinality from a single RGB Positioning in the image 3D Object is a long-standing problem in the field of computer vision . Driven by end-to-end deep learning , Recent research suggests that PnP Interpreted as a differentiable layer , such 2D-3D Point correspondences can be partially propagated through the back propagation gradient w.r.t. Learn from the attitude of objects . However , Start from scratch and learn a whole set of unlimited 2D-3D The point is difficult to converge under the existing method , Because the attitude of certainty is essentially non differentiable .

This paper presents a probabilistic algorithm for general end-to-end attitude estimation PnP layer ——EPro-PnP（end-to-end probabilistic PnP）, It's in SE Distribution of output attitude on manifold , Categorize substantively Softmax Bring into the continuous field .2D-3D Coordinates and corresponding weights are used as intermediate variables , By minimizing the relationship between the predicted attitude and the target attitude distribution KL Spread out to learn . Its basic principle unifies the existing methods , It's similar to the attention mechanism .EPro-PnP Performance is significantly better than other benchmarks , Reduced based on PnP Based on LineMOD 6DoF Attitude estimation and nuScenes 3D The gap between task specific methods of target detection benchmark .

EPro-PnP Method Overview .

Best student thesis nomination

This year's best student thesis was nominated by Harvard University and Google research 《Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields》 get .

• Institutions ： Harvard University, 、 Google research

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

Abstract of paper ： Neural radiation field is a popular view synthesis technology , It represents the scene as a continuous volume function , Parameterized by multilayer perceptron , The multi-layer perceptron provides the volume density of each location and the emitted radiation related to the view . Although based on NeRF The method is good at characterizing the appearance geometry of smooth changes , But they usually cannot accurately capture and reproduce the appearance of glossy surfaces . This study proposes Ref-NeRF To solve this problem , It will NeRF The parameterization of the emitted radiation associated with the view is replaced by the characterization of the reflected radiation , The function is constructed using a collection of spatially varying scene attributes . The study shows that , Use the regularizer on the normal vector , The new model significantly improves the authenticity and accuracy of specular reflection . Besides , The study also shows that the model can explain the internal characterization of the emitted radiation , This is very useful for scene editing .

And the best neural view synthesis model in the past mip-NeRF comparison ,Ref-NeRF Significantly improved the normal vector （ The top line ） And visual reality （ Residual downlink ）.

Other awards

Longuet-Higgins prize

Longuet-Higgins The prize is IEEE Computer association pattern analysis and machine intelligence （PAMI） The technical committee at the annual CVPR Issued by the 「 Computer vision foundation Contribution Award 」, In recognition of the great influence on computer vision research ten years ago CVPR The paper . The award is awarded to theoretical chemists and cognitive scientists H. Christopher Longuet-Higgins name .

This year's award-winning thesis is 2012 Published in 《Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite》, One of the three authors at that time Andreas Geiger and Philip Lenz Karlsruhr Institute of Technology , Raquel Urtasun From Toyota University of technology Chicago .

Picture source ： Twitter users @Kosta Derpanis

Address of thesis ：http://www.cvlibs.net/publications/Geiger2012CVPR.pdf

In this paper , The researchers used their own autopilot platform as a stereo 、 Optical flow 、 Visual range / SLAM and 3D Target detection and other tasks have developed a new and challenging benchmark . Their recording platform is equipped with 4 High resolution cameras 、1 platform Velodyne Laser scanner and 1 individual SOTA positioning system , Benchmarks include 389 Stereo and optical flow image pairs 、39.2km Long stereo range sequences and more than 20 m 3D Target notes （ Each image is visible at most 15 Cars and 30 Famous pedestrian ）.

The upper left is a recording platform equipped with sensors , The upper middle is the trajectory from the researcher's visual range benchmark 、 The top right is the parallax and optical flow diagram 、 Below is 3D Target tag .

Young Researcher Award

Young Researcher Award （Young Researcher Awards） In recognition of young scientists , encourage ta We continue to do pioneering work . The selection criteria is that the winners have less than years of doctoral degree 7 year .

The researchers who won the award this year are Bharath Hariharan and Olga Russakovsky.

Picture source ： Twitter users @Kosta Derpanis

Bharath Hariharan Assistant Professor, Department of computer science, Cornell University , Engaged in computer vision and machine learning , Especially those important problems that ignore big data labels .Hariharan The main research direction is to combine the progress of machine learning with computer vision 、 Geometry is combined with insights into specific areas of knowledge .

at present ,Hariharan My team is working on building a system , The system can understand tens of thousands of visual concepts with little or no supervision , Produce rich and detailed output , For example, the precise 3D shape , And reason about the world , Pass this reasoning on to humans .

He was in Google Scholar The number of citations of the papers on 25242,h Index is 38.

Personal home page ：http://home.bharathh.info/

Olga Russakovsky Assistant professor of computer science at Princeton University , We are committed to developing artificial intelligence systems that can reason about the visual world .Russakovsky The main research direction is computer vision 、 Human computer interaction and so on . Many of her papers were ECCV、CVPR Waiting to receive .

She is in Google Scholar The number of citations of the papers on 34756,h Index is 25.

Personal home page ：https://www.cs.princeton.edu/~olgarus/

Thomas S. Huang A Memorial Award

To commemorate the death of Thomas S. Huang（ Huang xutao ） professor ,PAMITC The Award Committee approved the establishment of Thomas S. Huang A Memorial Award , In recognition of CV Research 、 Researchers recognized as role models in education and services . The award comes from 2021 Issued since . Winners need to have at least a doctorate 7 year , It's best to be in the middle of your career development （ No more than 25 year ）.

Current Thomas S. Huang The winner of the Memorial Award is Professor lifeifei of Stanford University .

Picture source ： Twitter users @Kosta Derpanis

Lifeifei is a professor of computer science at Stanford University , NAE member , Academician of the National Academy of Medicine . Her major fields are computer vision and cognitive neuroscience .2016 year , Li Feifei joins Google China Center team of cloud AI and machine learning , With Google Cloud As the chief scientist, he is one of the team leaders .2018 year 9 month , She announced her return to teach at Stanford University , And continue to participate in Stanford University AI Topic research .

Lifeifei's work includes cognitive - inspired AI, machine learning , Deep learning , Computer vision and AI + Health care , In particular, environmental intelligence systems for healthcare delivery . She also works in cognitive and computational neuroscience . She invented ImageNet and ImageNet Challenge, among ImageNet Challenge It is an important large-scale data set and benchmark work .

She is in Google Scholar The number of citations of the papers on 167561,h Index is 128.

Personal home page ：https://profiles.stanford.edu/fei-fei-li

 THE END 

 Past highlights 




 It is suitable for beginners to download the route and materials of artificial intelligence ( Image & Text + video ) Introduction to machine learning series download Chinese University Courses 《 machine learning 》（ Huang haiguang keynote speaker ） Print materials such as machine learning and in-depth learning notes 《 Statistical learning method 》 Code reproduction album machine learning communication qq Group 955171419, Please scan the code to join wechat group