Click on the “CVer“, select “Star” or “Top”.
Important content delivered promptly.
Editor: Amusi Source: Zhihu
https://www.zhihu.com/question/353691411
This article is for academic sharing only. If there is an infringement, the article will be deleted.
How to Derive Your Ideas from Others’ Papers in Computer Vision?
Author:Cheng Lihttps://www.zhihu.com/question/353691411/answer/900046621
I’ve noticed that many people are reading this (to avoid misleading), I feel I need to add some explanations.
If it’s just A+B, you need to explain why it’s non-trivial.
Generally, it should at least be modified to A+B’
Or A+B+C is easier to publish.
For example, I previously worked on a paper.
It was actually Unsupervised Landmark + VUNet’s decompose + CycleGAN.
If I had just tried to submit the first two terms halfway through to ICLR, it wouldn’t have worked…
Later, after completing the CycleGAN part, I got accepted at CVPR…
(Since it seems like many people liked it, we will organize our publications later.Actually, many times A+B’ can also lead to some pretty good ideas.)
Original Answer:
I actually have a good idea…
Find 40 relatively new oral papers.
Preferably open-source, understandable, as trendy as possible, and liked by experts.
Then draw a 40*40 matrix…
Do not consider the diagonal elements, leaving 1560 elements.
Check if A+B is reliable for each element.
Although 99% of them may not be reliable…
However, it is still possible to filter out around 15 ideas…
(If considering commutativity, even 7 papers would be enough…)
Or you can find 40 relatively new oral papers that are not yours,
Then find K of your own papers, and you can do this as well.
This way, you don’t have to exclude the diagonal elements.
My personal publication level is still not high.
However, many are not generated by A+B…
For example, regarding CNN before,
Some were about publishing datasets.
Pixel-level hand detection in ego-centric videos.
Some actually involve many steps in a classic pipeline with A+B+C.
Others discuss steps B and C more, but step A is also very important.
Coming up with a trick for A eventually led to a paper.
Face alignment by coarse-to-fine shape searching.
A+B can also produce interesting ideas when the span is large…
It’s not simply incremental work.
For example, applying recommendation systems to classifier recommendations (before the CNN era)…
Model recommendation with virtual probes for egocentric hand detection.
Divide and conquer is also a common approach; any topic can be added (before the CNN era).
Unconstrained face alignment via cascaded compositional learning.
Sometimes, seeing others’ RL+tracking papers inspires ideas for clustering methods, leading to a new paper.
A+B (although the steps were a bit large, it was often rejected and later submitted to AAAI).
Merge or not? learning to group faces via imitation learning.
This year, I also saw someone using GCN for clustering, so I combined it with GCN and resubmitted a paper…
(Not released yet)
Similar to this…
Sometimes, you can also engage in philosophical discussions, which aren’t just simple A+B.
The devil of face recognition is in the noise.
I’m worried that my bosses will see this and say I’m misleading people…
Author:Cheng Xuyuan, who writes bugshttps://www.zhihu.com/question/353691411/answer/897499123
Ideas cannot be generated solely by reading papers; many scenarios may yield ideas, and this requires inspiration.
You should mean research scope or research problems, right?
Generally, start with the introduction and conclusion to understand what the paper is about, its contributions, experimental results, and future work prospects. Future work can serve as an inspiration. If the research direction and algorithms used in the article interest you, check the experiments for design ideas and frameworks, and look at the discussion for experimental analysis; areas where results were poor may also be research problems.
Corresponding solutions—your ideas—may involve your theoretical foundation, inspiration during paper replication, practical application scenarios, and more.
If the above is too abstract, here’s a simple statement: replicate the papers you are interested in, and think more during the replication process to generate ideas.
Author:Zhang Xiaoyuhttps://www.zhihu.com/question/353691411/answer/899997687
Start with the paper’s title, abstract, introduction, conclusion, and discussion.
First, discuss the title.
Every year, there are many CS papers, but we don’t have the energy to read them all, so we can filter out many directions we don’t care about through the title, reducing the energy cost of finding ideas.
Next, start with the abstract to understand what problem the paper mainly addresses, what the general method is, and what the final conclusion is. Firmly grasp these three points.
Finally, look at the discussion and conclusion sections; these are often key areas for finding ideas. In the discussion section, identify which issues are unclear or inadequately addressed, or where the reasons for certain phenomena are not thoroughly explored; these aspects can be attempted to be mined. The same goes for the conclusion section: see what methods and evaluation metrics the author used, what conclusions were drawn, and consider whether the methods are optimal, whether the evaluation metrics are the best, and if changing them would yield consistent conclusions. If consistent, this can validate the paper; if not, where does the discrepancy lie? What factors do we need to consider in this type of research?
Additionally, organizing and categorizing the papers is also very important. After reading a certain number, you will understand the main research methods for a specific problem, how well they are developed, and where understanding needs improvement, deepening, or filling gaps. This can lead to innovative methods in other fields or even proposing innovative methods, which is all part of daily accumulation.
Author:LeapMayhttps://www.zhihu.com/question/353691411/answer/910600098
First, reproduce the code from others’ papers to understand the methodology and code, and thoroughly comprehend their papers.
Second, carefully consider the conclusions and future work directions in the conclusion section of the paper.
Third, identify the shortcomings in the methods of the paper; can the model be optimized? Can the ideas be simplified? Combine this with your own knowledge to preliminarily judge the points you can explore, and then try to discuss and communicate with experts around you.
Fourth, based on the points you can focus on, start theoretical derivation and engineering implementation to further refine your ideas.
Important! The CVer academic exchange group has been established.
Scan the code to add the CVer assistant and apply to join the CVer large group and specialized direction groups.The specialized directions covered include:Object Detection, Image Segmentation, Object Tracking, Face Detection & Recognition, OCR, Pose Estimation, Super Resolution, SLAM, Medical Imaging, Re-ID, GAN, NAS, Depth Estimation, Autonomous Driving, Reinforcement Learning, Lane Detection, Model Pruning & Compression, Denoising, Dehazing, Deraining, Style Transfer, Remote Sensing Images, Action Recognition, Video Understanding, Image Fusion, Image Retrievaletc.Make sure to note:Research Direction + Location + School/Company + Nickname(e.g., Object Detection + Shanghai + SJTU + Kaka), following this format will expedite approval and invitations to the group.
▲ Long press to join the group.
▲ Long press to follow us.
Please give me a thumbs up!