Reference: https: //pan.baidu.com/s/1MfcmXKopna3aLZHkD3iL3w
First, why should read the papers?
- Basic technologies: Read the papers related works can help us understand some of the key technologies in the field, the field of classical algorithm to track the papers
- New directions and ideas: learn about the latest thinking in the field to solve
- Reproduction: a more thorough understanding of the algorithm logic, coding exercise capacity
Second, what papers to read?
- Summary: Summary contains the classic direction of the development process, to help replenish the necessary basic knowledge, familiar with some of the key algorithms and technology stack, you can verify the effect in practice. Quickly find algorithmic solutions for the scene
- Representative papers: the most critical technical papers in the field (high-reference), movable from review, usually by the open source code that can help to quickly verify ideas.
- Frontier thesis: Learn about the latest developments in the field of ideas, to learn new ideas in the process of landing in products, innovation and efficiency to help solve the problem.
Third, where to find the paper?
- Top will: CVPR, AAAI, ICCV etc.
- google scholar: https://scholar.google.com/
-
Watch VALSE & Webinar: VALSE China is a large-scale symposium meetings, held every year to explain the algorithms and papers from various quarters such as the annual progress; video algorithms usually have to explain;
-
Papers with Code: https://paperswithcode.com, finishing a large number of papers, and reproduce the original author or the code; convenience while watching the paper side of the experiment;
Fourth, how to read the paper?
-
Origins: Sources have some introductory, such as paper Google Brains, FAIR and other large research institutions are usually easy to read, easy to reproduce, or have open source code can practice;
- Abstract: looking to see whether the direction of the main innovations of interest, and whether, to avoid wasting time; often referred to innovation and effect (accuracy, AP, etc.);
-
Introduction: fast reading, retelling the main problem in comparison algorithm (to help you know what problems exist before or algorithm being used) and the proposed innovation;
-
Related Work: If you do not understand the correlation algorithm classic, you can find out read; If you already know, you can choose to skip;
-
Experiments: the accumulation of experience training algorithms and data preparation; look at the results on which the leading indicators;
Fifth, how "writing papers"?
-
Read more: take full advantage of cutting-edge ideas and summarize the problem; there is no perfect algorithm, only better ideas;
-
Taki: read the paper side edge annotation own thinking and ideas in the next, and the similarities and differences on other papers and seen the idea;
-
Write: classic papers referring to a common format, written in his own experiments to do a "paper"; record ideas and experimental results;
-
More discussion: discuss their ideas and experimental methods with the same group or Daniel, lessons learned and issues;