Full reappearance He Cai Ming ICCV award-winning thesis results and open source! (attached thesis & Open source code)

Sorted from: AI Technology Review, AI technology base camp, Tucson future

This article length is 1434 words, suggest reading 3 minutes

This article introduces you to the 20,172 award-winning papers of He Cai Ming ICCV, and enclose the open source code to reproduce the results of the thesis.

ICCV, the top conference on computer vision, received 2,143 papers in 2017, up 26.2% from 1698 in the previous ICCV2015. A total of 621 articles were selected as the General Assembly papers, employing the proportion of 28.9%;poster, spotlight, oral ratio of 24.61%, 2.61% and 2.09% respectively.

According to the author's signature, the organizing committee has counted the number of research at different universities, Tsinghua University, CMU, MIT, ICL, Stanford and Google ranked first, Shanghai Jiao Tong University and Beijing Beihang College are also among the top 10.

And the biggest news of this conference must be He Cai the great God in ICCV to win the double best paper. After two times won the CVPR Best thesis Award, He Cai participated in the two most recent papers, respectively, ICCV 2017 of the Best essay Award (top Paper Award) and the best student thesis (Top Student Paper Award) two awards.

The two winning papers, respectively, were released this year in April, "Mask r-cnn" and the "Focal Loss for dense Object detection", released this August, and are only 4 months apart. Be aware that ICCV is one of the top conferences in the field of computer vision and is held twice a year, and He Cai as the first and fourth author of two papers to prove his strength.

Introduction of two winning papers:

Suorce: Know

Abstract: We propose a simple, flexible and Universal object instance segmentation framework. Our method can effectively detect objects in images and generate high-quality segmentation mask for each instance. This method, called Mask r-cnn, extends faster r-cnn by adding a branch for the Prediction object Mask, which is parallel to the existing branch that is used for bounding box recognition. Mask R-CNN Training is simple, just add a smaller overhead on top of the faster r-cnn running in 5fps. In addition, Mask R-CNN can easily be extended to other tasks, such as allowing attitude estimation in the same frame. We have achieved the best results in the three orbital tasks of the COCO series, including instance segmentation, boundary object detection and human key point detection. No tricks,mask R-CNN has outperformed all existing single models, including the COCO 2016 Challenge champion.

Paper Address:

https://arxiv.org/abs/1703.06870

Suorce: Know

Abstract: The most accurate target detector is a two-stage method commonly used in R-CNN, which applies classifiers to a sparse set of samples composed of candidate target locations. Instead, the one-stage detector is applied to a regular, dense set of samples made up of possible target locations, and is faster and simpler, but the accuracy lags behind the two-stage detector. In this article, we discuss the cause of this phenomenon.

We find that the major cause of this phenomenon is the severe foreground-background category imbalance that occurs during the training of dense target detectors. Our solution to this type of imbalance (class imbalance) is to reshape the standard cross entropy loss so that it reduces the weight of the loss of the classified sample. The Focal Loss concentrates the training on a sparse set of difficult samples and prevents large numbers of simple negative samples from drowning the detectors during the training process. To assess the effectiveness of this loss, we designed and trained a simple dense target detector-retinanet. The test results show that when using Focal loss training, the retinanet not only catches the one-stage detector's detection speed, but also surpasses all the current state-of-the-art two-stage detectors in accuracy.

We propose a new loss function focal Loss (focus loss), which adds a factor (1-pt) γ to the standard cross entropy standard. Setting γ> 0 reduces the relative loss of a well-defined sample (pt >. 5), making the model more focused on samples of difficult error classifications. The results show that with a large number of simple background samples (background example), our proposed Focal Loss function can train the dense object detector with high accuracy.

Paper Address

https://arxiv.org/abs/1708.02002

People have expressed their admiration for the great God of Kai-ming and the best academic researchers. At the same time, many practitioners are more concerned with the question of when to see open source code.

"The best way to honor the great God, perhaps, is to reproduce the results of the paper and then open it again," says Tucson, an AI tech company. They fully reproduced the results of He Cai's thesis (Mask r-cnn and Feature Pyramid Network) and made the corresponding code open source. This is also the first open source code that can fully reproduce the results of He Cai's thesis.

Backstage reply the keyword "open source", obtains the complete reappearance He Cai The result in the paper the Open source code, hoped can give everybody the work to bring the help.

The He Dashen of gossip and the Hanging of life

The ICCV won the two theses awards not He Cai the first time the great God demonstrated his ability to open up.

2003, He Cai Ming to the Guangdong Provincial College entrance examination as the top scholar to enter Tsinghua University.

In 2009, at the IEEE CVPR Conference, the first paper "single Image Haze removal Using Dark Channel Prior" (MSRA) internship at the Microsoft Institute of Asian Studies (He Cai) was first fully A team made up of Chinese people won the award.

In 2016, He Cai another paper from the team "Deep residual Learning for Image recognition" and won the CVPR Award for best thesis.

In August, He Cai left Msra and joined FAIR (Facebook AI) as a scientist.

After joining Facebook, He Cai soon became a major contributor to faster R-CNN and Mask r-cnn, and there was no adaptation period.

Now, He Cai also become ICCV's best thesis winner, once again with me and other ordinary people's gap, people can not help.

However, this is not the most desperate, take a look at He Cai-ming Weibo, travel, play games ... Everything is not falling down. So the great God is the great God after all ...

Small knitting can only silently kneel to the great God ...

You are welcome to leave a message to exchange feelings