[academic related] how to find the innovation of top papers? Chinese universities won the CVPR Best Student Thesis Award for the first time

A few days ago , Tongji University announced a big news ： Our student chenhansheng won CVPR2022 Best student thesis Award , This is also CVPR since 2001 Since the establishment of the best student thesis award in , The first author of the award-winning thesis is from a Chinese University for the first time .

Chinese in CV The rise of the field

Recent years CV Top of the field , Chinese writers are increasing day by day . With CVPR 2022 For example , The authors from Chinese Mainland contributed the most , Accounted for 44.59%; The United States uses 20.65% In the second place , Less than half of the first place . Although our insiders have been used to the soaring of Chinese at the summit , However, students in Chinese Mainland won such a blockbuster award as the best student thesis , It's still very shocking ！

What is this masterpiece for ？

I read the title of the author's paper ：EPro-PnP: Generalized End-to-End Probabilistic Perspective-n-Points for Monocular Object Pose Estimation. Suddenly feel familiar and strange , Admire with a little stunned ,PnP Such a traditional algorithm is still being studied and improved silently , And made breakthrough work , Recognized by peers all over the world . therefore , I downloaded the paper with great interest , I want to see if I can learn something .

This award-winning paper mainly studies objects in monocular images 6 Pose estimation of degrees of freedom , Proposed EPro-PnP It's an end-to-end probability PnP Algorithm , It combines traditional multi view geometry with deep learning network , Achieved in a single RGB Accurate pose estimation on the image , This method has achieved excellent results in public data sets of pose estimation and 3D object detection . It can be widely used in the field of robots and automatic driving . The author also open source code ：

https://github.com/tjiiv-cprg/EPro-PnP

This method is a general end-to-end 2D-3D Corresponding solutions , This “ Universal ” It's very important , How important is the generalization of the model , Everyone who knows knows .

Inspiration about innovation

In us SLAM Knowledge of the planet Communication community , One of the most anxious things for my friends is the innovation of the paper . I previously made a live broadcast about the innovation of the paper ： Tonight? 19 Click live | Do research on how to find innovation ？  The core points are as follows , I also gave an example to lead you to find innovative points , If you are interested, you can go to the planet to watch the playback video .

1、 Read all the review papers in the big field ：survey/review + key word
2、 Choose the subdivision direction according to the accumulation and interest
3、 Read the subdivision direction recently 3-5 Review papers in 、 Chinese dissertation 、 English top conference paper
4、 Pay attention to papers with open source code 、 Compile operation ; Try to reproduce a particularly good paper
5、 Collect data by yourself 、 Hardware transplantation 、 test ; If you have the ability , Refer to previous experience and write a framework by yourself
6、 Problems found in the actual project . Check the literature to see how others solved it ; Think about how to solve
7、 start to do ！

Read this article EPro-PnP The paper of , Let me add a few more points ：

1、 Pay attention to the basic knowledge in the field of computer vision .

as far as I am concerned , Many other majors have switched to CV Generally, the direction is to study deep neural networks directly , They may be familiar with all kinds of classic Networks , Finally, he became a scalper 、 Diao Shen Xia .

But it ignores CV The importance of basic knowledge , For example, the most basic histogram equalization 、 Morphological opening and closing operation 、 Camera imaging model 、 Polar geometry 、PnP、ICP etc. . And sometimes , These basic knowledge determine your upper limit .

Here's the thing to note , The text in the above figure is PS Of , The irony is even greater , We can use homography matrix in multi view geometry to realize the above P Picture effect . If you know how to do it, you can comment in the message area .

2、 Traditional multi view geometry and Deep neural network （DNN） The combination of .

Now? DNN The research direction has gradually shifted from two-dimensional to three-dimensional , If you are familiar with the core knowledge of multi view geometry in 3D vision , It can definitely make your research work even stronger . Multi view geometry has been studied for many years , Theoretical basis with certainty , It can describe the structure of the physical world 、 shape 、 Volume 、 depth 、 Posture 、 parallax 、 motion 、 Optical flow and other information , Very interesting , It is essential in 3D vision . If we can combine the two , It is very promising to make some breakthroughs .

The above is my personal nonsense , If it works for you , Remember to watch one by one ~

 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