Click “Zizhi” to get more AI knowledge!

The International Conference on Computer Vision (ICCV 2017), a biennial International Conference on Computer Vision (ICCV 2017), is being held in Venice, Italy, from October 22 to 29. Computer vision experts from around the world gathered in Venice to present the latest advances in computer vision and related fields. Awards announced include the Marr Prize, Best Student Paper, Honorable mentions, Azriel Rosenfeld Lifetime Achievement Award, and Distinguished mentions Researcher Award, Everingham Prize, Helmholtz Prize. Facebook is the big winner. The Marr Prize and the Best Student Paper went to Kaiming He, a research scientist at Facebook’s ARTIFICIAL intelligence lab. Jia Yangqing Caffe team won the Everingham Prize.

ICCV 2017 profiles

The International Conference on Computer Vision (ICCV), the Conference on Computer Vision and Pattern Recognition (CVPR), and the European Conference on Computer Vision (ECCV) are the top three conferences in the field of computer vision. This conference covers 3d stereo vision, medical image analysis, face and pose learning, low-level vision and image, motion and tracking, target detection and recognition, optimization methods, image segmentation and edge extraction, statistical methods and learning, video event detection and behavior recognition and other fields. Since its inception in 1987, the conference has been held every two years, with the practice of rotating in all continents of the world. The active participation of experts in various fields and industry has promoted the process of computer vision technology from germination to development, from laboratory prototype to practical application.

In terms of the number of participants, the number of participants this year was 3,107, breaking the 3,000 mark, more than double the number of participants in the last ICCV 2015, which shows how hot the field of computer vision has been in recent years.

In terms of papers, the number of admitted papers is the largest in history. This conference received 2,143 papers from all over the world and 621 papers were accepted, including 43 oral reports and 520 poster papers. The acceptance rates were 2.09% and 28.9% respectively, showing the fierce competition.

ICCV 2017 major awards

Marr Prize

,Article:Mask R-CNN“

  • Author: Kaiming He, Georgia Gkioxari, Piotr Dollar, Ross Girshick

  • The paper address: https://research.fb.com/publications/mask-r-cnn/

  • Arxiv：https://arxiv.org/abs/1703.06870

  • This paper proposes a compact and flexible universal object instance segmentation framework. Mask R-CNN can not only detect the target in the image, but also give a high quality segmentation result for each target. It extends the Faster R-CNN and adds a new branch for objectmask prediction on the bounding box recognition branch in parallel. The network can also be easily extended to other tasks, such as estimating a person’s posture, known as Person Keypoint detection. The framework achieved the best results in a series of COCO challenge tasks, It includes instance segmentation, bounding-box object detection and Personkeypoint detection.

 

    

Best Student Paper award

,Paper:“Focal Loss for Dense Object Detection“

  • Author: Tsung-Yi Lin, Priya Goyal, Ross Girshick,Kaiming He, Piotr Dollar

  • The paper address: https://arxiv.org/abs/1708.02002

  • Existing object detection methods based on deep learning are mainly divided into two categories, one is two-stage (represented by ftP-RCNN) and the other is one-stage (represented by YOLO and SSD). The current one-stage method has the advantages of fast speed and simple model compared with two-stage method, but its accuracy is lower than two-stage method. In this article, the author found that the main reason for this phenomenon is the extreme imbalance in the proportion of positive and negative samples. In this paper, a new loss function is constructed, which can dynamically adjust the cross entropy and solve the problem of sample imbalance through the influence of difficult and easy samples on the loss value. The simple and significant experimental effect can be achieved due to the previous online difficult case mining (OHEM) and other methods, which can greatly improve the performance of one-stage object detection method.

Other awards:

N Rhinehart, KM Kitani

First-Person Forecasting with Online Inverse Reinforcement Learning

Tomaso Poggio

 

Home page: https://mcgovern.mit.edu/principal-investigators/tomaso-poggio

 

Tomaso Poggio, 67, is the Eugene McDermott Chair Professor in the Department of Brain and Cognitive Sciences at MIT’s Artificial Intelligence Lab and co-director of the school’s Center for Biology and Computer Learning. Poggio began teaching at MIT in 1981, after spending a decade in the Biocybernetics lab at the Max Planck Institute in Tubingen, Germany. He received his doctorate from the University of Genoa in 1970. Poggio is a foreign member of the Italian Academy of Sciences and a fellow of the American Academy of Arts and Sciences.

 

Luc Van Gool

Home page:

https://www.vision.ee.ethz.ch/en/members/get_member.cgi?id=1

 

Professor at the University of Leuven in Belgium, co-founder and chief technical Advisor of Kooab, an image recognition company that was later acquired by Qualcomm. Prof. Dr. Gool is also co-founder of eSaturnus, Eyetronics, GeoAutomation and Procedural Inc. He also serves as editor-in-chief of the Journal Foundations & Trends in Computer Graphics and ComputerVision. He heads the computer vision Laboratory at ETH Zurich in Switzerland and is a professor at the University of Leuven in Belgium. His main research interests are 3D reconstruction and modeling, target recognition, visual tracking, and robot control.

 

  

Richard Szeliski

Home page:

http://szeliski.org/RichardSzeliski.htm

Dr. Richard Szeliski is a master in the field of computer vision. Dr. Szeliski has more than 25 years of experience in computer vision research at Digital And Microsoft Research. In 1996, during his tenure in Microsoft Research Institute, he proposed a motion based panoramic image Mosaic model, which uses L-M algorithm to match images by finding geometric transformation relations between images. This method is a classic algorithm in the field of image Mosaic, and RichardSzeliski became the founder of the field of image Mosaic.

Yangqing Jia    http://daggerfs.com/

The great God He Kaiming

In addition to this ICCV2017, Dr. He Kaiming won the best paper award and the Best Student Paper Award. On Facebook, Cool!

Personal academic homepage: http://kaiminghe.com/

In addition, he has been awarded CVPR Best Paper Award twice in CVPR 2016 and 2009. To recap:

CVPR2016 Best Paper Award

Paper:   “Deep Residual Learning for Image Recognition“

  • Author: Kaiming He, XiangyuZhang, Shaoqing Ren, Jian Sun

  • The paper address: https://arxiv.org/abs/1512.03385

  • This paper proposes a framework for residual learning to reduce the training burden of the network, which is much deeper than previous networks, and explicitly uses layers as input to learn residual functions, rather than unknown functions, in response to the fact that deeper neural networks are often more difficult to train. This paper provides very comprehensive experimental data to prove that residual networks are easier to optimize and can increase accuracy as depth increases. A residual network with a depth of 152 layers (8 times that of VGG) is evaluated on ImageNet’s dataset, but still has a lower complexity than VGG. The residual network achieved an overall error rate of 3.57%, which won the first place in the classification task of ILSVRC2015. Cifar-10 data set was also used to analyze the 100-layer and 1000-layer networks. Based on the deep residual network, this network made the submitted version to participate in ILSVRC and COCO2015 competition, and won the first prize of ImageNet object detection, ImageNet object localization, COCO object detection and COCO image segmentation.

CVPR Best Paper Award 2009

       Paper: “Single Image Haze RemovalUsing Dark Channel Prior“

  • Author: Kaiming He, Jian Sun,Xiaoou Tang

  • The paper address: http://ieeexplore.ieee.org/document/5206515

  • This paper studies the problem of image defogging technology, which can restore the color and visibility of images, and estimate the distance of objects by using the concentration of fog. All of these have important applications in computer vision (such as 3D reconstruction and object recognition). But there hasn’t been a simple and effective way to do this. In this paper, the paper found a very simple, even surprising statistical law, and proposed an effective method to remove fog. Different from the previous methods, this paper focuses on the statistical characteristics of fog-free images, and finds that in fog-free images, every local area is likely to have shadows, or pure color things, or black things. Therefore, every local region is likely to have at least one color channel with a very low value. Call this statistical rule Dark Channel Prior. Intuitively, Dark Channel Prior believes that there is always something Dark in every local area. This rule is very simple, but it is the essential basic rule in the fog removal problem we study. The proposed Dark Channel Prior can effectively remove the influence of fog and estimate the distance of objects by using the concentration of objects.