Writting the Paper - 程明明个人主页

46
Theory Algorithms Experiments Conclusions Acknowledgements References Appendices Writting the Paper Theory, algorithms, experiments, conclusions, etc. Ming-Ming Cheng College of Computer Science 2020/05/08 Ming-Ming Cheng Nankai University The paper 2020/05/08 1 / 46

Transcript of Writting the Paper - 程明明个人主页

Page 1: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Writting the PaperTheory, algorithms, experiments, conclusions, etc.

Ming-Ming Cheng

College of Computer Science

2020/05/08

Ming-Ming Cheng Nankai University The paper 2020/05/08 1 / 46

Page 2: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Table of Contents

1 Theory

2 Algorithms

3 Experimental Results

4 Conclusions

5 Acknowledgements

6 References

7 Appendices

Ming-Ming Cheng Nankai University The paper 2020/05/08 2 / 46

Page 3: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Fig.: Sections 3.6-3.10.

Ming-Ming Cheng Nankai University The paper 2020/05/08 3 / 46

Page 4: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Theory

A formal statement of the theorem or result itself.Describle how the results fits into the overall solution.

A sub-problem goal that is addressed by the result.

A proof or other reasoning to validate the results.

Be readable after remove equations!

Ming-Ming Cheng Nankai University The paper 2020/05/08 4 / 46

Page 5: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Theory: explain with examples

Some people may learn well from theoretical arguments,but aothers learn better from examples.

Helpful to give some simple examples to show howthe theory works in practice.Give readers confidence that they understand youridea, and the idea actually works.

Not just equations, but also explain what your approach isWhat the euqation mean.Why the equation take the form they do.

Ming-Ming Cheng Nankai University The paper 2020/05/08 5 / 46

Page 6: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Theory: Example

Fig.: A counterexample: (a) input image containing four regions andfour seeds of three types; (b) random walk segmentation result;probability of a random walk starting at each pixel first reachingseed type (c) sU or (d) sO .

A property suggested by [Grady 06] is that each segmented re-gion is guaranteed to be connected to one or more seeds withthat region’s label: Isolated regions without seeds do not occur.

Connectedness of Random Walk Segmentation, IEEE TPAMI 2011.

Ming-Ming Cheng Nankai University The paper 2020/05/08 6 / 46

Page 7: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Theory: ExampleTo show this, we analyze the random walk problem usingan equivalent electrical circuit network. We further assumethat the algorithm works on general graphs, and generalizethe segmentation problem to a special convex hull partitionproblem (Observ. 1). Then we show, step by step (Propos.1-5), that for any segmented graph which satisfies the con-nectedness property and has more than three label types,even allowing the algorithm to have more general weight-ing (Propos. 4) and assignment rule (Propos. 5), it is stillpossible to replace some part of it in a way which makesthe connectedness property fail. Finally, Proposition 5 pro-vides a condition under which random walk segmentationis guaranteed to give a connected result.Connectedness of Random Walk Segmentation, IEEE TPAMI 2011.

Ming-Ming Cheng Nankai University The paper 2020/05/08 7 / 46

Page 8: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Theory: Example

Ming-Ming Cheng Nankai University The paper 2020/05/08 8 / 46

Page 9: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Theory: Example

Fig.: An example of counterexample construction: initial seeds areshown in color in (a) and the determined image colors and finalpixel labels are shown in (c). (a) Graph and seeds. (b) Coun-terexample. (c) Image color.

Ming-Ming Cheng Nankai University The paper 2020/05/08 9 / 46

Page 10: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Algorithms: examples

How does it contribute to the overall solution?

This algorithm sharpens the input image, 2whichimproves the reliability of the text detection process.

What are the inputs and assumptions?

This algorithm has three inputs. These are two images,which should have the same dimensions, no greater than1024 × 768 pixles, and a user-selected mixing ratio r inthe range [0,1].

What happens in exceptional cases?

If two copies of the same image are provided as input, · · ·

Ming-Ming Cheng Nankai University The paper 2020/05/08 10 / 46

Page 11: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Algorithms: examples

What are the outputs?

The output is a new image with the same dimension asthe input image. A fraction r of its pixels are chosen atrondom from the first image, · · ·

For iterative algorithm, how to initialize?

The car tracking aorithm iteratively seeks the position ofthe car. In each video frame, this iteration is started by · · ·

Can user control the results, speed/accuracy trade-off?

The user must also supply a real tolerance t whichcontrols the accuracy of the results · · ·

Ming-Ming Cheng Nankai University The paper 2020/05/08 11 / 46

Page 12: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Algorithms: state in correct logical order

Do not write like this!

Read in the input data and process it to give the output,removing any noise from the input if necessary.

Explain the procedure in the correct logical order.

Read in the input data, remove any noise from it ifnecessary, and process it to give the output.

Ming-Ming Cheng Nankai University The paper 2020/05/08 12 / 46

Page 13: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Algorithms: long descriptions is hard to follow!

Especially if there isNested loopingOr complex branching.

Use pseudocode listing!

Ming-Ming Cheng Nankai University The paper 2020/05/08 13 / 46

Page 14: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Algorithms: main program then each proced.Describe the main program, briefly describing thepurpose of any procedures it calls. Subsequentlydescrible each procedure in more detail, as necessary.

Algorithm 1 Text Summarization Algorithm

1: procedure TEXT DOCUMENT SUMMARIZATIONInput: A text document.Output: Summary sentences.

2: Creating information table from input document.3: Generate matrices.4: Call Procedure: Reduct construction [Alg. 2]5: Return: Summary sentences6: end procedure

Ming-Ming Cheng Nankai University The paper 2020/05/08 14 / 46

Page 15: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Algorithms: explain the pseudocode

Describle how the pseudocode works.Explain individual lines,or group of lines.

Text to explain, not repeat it in words!

Ming-Ming Cheng Nankai University The paper 2020/05/08 15 / 46

Page 16: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Algorithms: tuning parameters

It is important to disscuss all the tuning parameters.Suggest default/typical values and the range.Explain how you chose such values and range.Given a forward reference to experiments section ifthe values were determined experimentally.

Justify all tuning parameters!

Ming-Ming Cheng Nankai University The paper 2020/05/08 16 / 46

Page 17: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Algorithms: how stable is the algorithm?

Experimentally verify how stable the algorithm is w.r.t.changes in the parameters.

Much more convincing if you use same parametersfor all your examples, rather than pick for each input.Fewer parameters means easier to use. Too muchcontrol freedom would be confusing.To what extend is the algorithm automatic?Make clear in all places where user input is required,and explain what form that user input takes.Explain how to choose suitable input.

Ming-Ming Cheng Nankai University The paper 2020/05/08 17 / 46

Page 18: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experimental Results

For technical papers, this can be benchmark results.For theoretical papers, experiments could testaggreement between predicted values and thoseobtained by observation.Reports what you did, and what you found out, thusshould be written in the past tense. The major part ofthe paper should be written in the present tense.

Ming-Ming Cheng Nankai University The paper 2020/05/08 18 / 46

Page 19: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: each have a seperate purpose

Each should have a seperate, identified purpose.

Results for different kinds of input data.

Show the algorithm results w.r.t different kinds of inputdata, both typical, and extreme or unexpected.

Choosing parameters.

Show what happens with different settings of parameters,and how to optimally choose such parameters.

Not just be a set of results.

Ming-Ming Cheng Nankai University The paper 2020/05/08 19 / 46

Page 20: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: each have a seperate purposeEach should have a seperate, identified purpose.

An experiments could validate some aspect of a theory.

An experiment could provide a measure of the quality ofthe results of an approach.

It is important to perform tests which enable comparisionswith existing methods or theories wherever feasible.

To get published, the idea needs to be better in some way.

Ming-Ming Cheng Nankai University The paper 2020/05/08 20 / 46

Page 21: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: objective if possible

You should give objective, numerically quantified, resultswhenever possible.

Our approach to image enhancement takes 7 s toprocess an image of size 1024 × 1024 pixels.

Figure 2 presents the precision recall curve of thesaliency maps for thresholding in the range [0, 255].

Ming-Ming Cheng Nankai University The paper 2020/05/08 21 / 46

Page 22: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: objective if possible

First define metrics which will be used to evaluateyour works against others.Collect the scores for each according to the chosenmetrics.Be careful to make fair comparisions by running thetest in an unbiased way.Papers only with qualitative, and rely on subjectivejudgement is rather hard to get accepted.

Ming-Ming Cheng Nankai University The paper 2020/05/08 22 / 46

Page 23: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: stepsClearly stating its purpose, then how the experimentwas performed.Give the results of the experiments.Provide an interpretation of the results, makingexplicit what you infer from them.Explain how and where you obtained any data usedfor testing, and why this particular dataset was used.Use standard benchmark with GT data if possible.Be honest in your descriptions of your experiments.All methods have limitations.Make sure that any conclusions you draw aresupported by the results of your experiments, andthat you do not over-generalise.

Ming-Ming Cheng Nankai University The paper 2020/05/08 23 / 46

Page 24: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: state the purpose

Clearly stating its purpose and how the experiment wasperformed.

The aim of this experiment was to determine the radial lensdistortion in the camera lens used in our system. An imagewas taken of the reference target describled in Section 3.2.and the first and second order radial · · · using Algorithm 2and Equation 6.

Ming-Ming Cheng Nankai University The paper 2020/05/08 24 / 46

Page 25: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: results

Give the results of the experiments.

The first and second order radial and tangential distor-tion coefficients were determined to be K1 = 0.00037m−2,· · · ,P2 = 0.0000017m−4.

Ming-Ming Cheng Nankai University The paper 2020/05/08 25 / 46

Page 26: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: interpretationProvide an interpretation of the results, making explicitwhat you infer from them.

This might be how well your approach works in compar-ision to other methods, or how an algoirhtm scales withquantity of input data, or what the optimal conditions are.

Using Equation 9, the magnitude of these lens distortioncoefficients indicate that, without correcting for lens dis-torition, the position of the robot can only be determinedto within 8.5 cm, which is not adequte for safty. It is thusnecessary to correct for lens distortion.

Ming-Ming Cheng Nankai University The paper 2020/05/08 26 / 46

Page 27: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: explain the data

Explain how and where you obtained any data used fortesting, and why this particular dataset was used.

To test our algorithm for removing motion blur, we collectedas test cases the first 50 images returned by searchingGoogle Image Search using keywords ’motion blur image’.These exhibited a range of subjects, and differing amountsof motion blur, allowing us to test how well our algorithmworks with varying inputs.

Ming-Ming Cheng Nankai University The paper 2020/05/08 27 / 46

Page 28: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: benchmark with ground truth

Use standard benchmark with GT data if possible.

For example, the ImageNet dataset have been widely usedin image classifcation, and come with ground truth. TheMS COCO dataset have been used in many papers onobject detection, and come with ground truth.

Crucial for fair comparisions!

Ming-Ming Cheng Nankai University The paper 2020/05/08 28 / 46

Page 29: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: honest descriptions

Be honest in your descriptions of your experiments. Clearlyexplain:

What was successful and worked well.What did not work as expected, explain why not.Giving examples which show your method/theoryworks well, on the margin, and breaking down,allowing readers to better understand andcharacterise your achievements.

Avoid biasely selected examples.

Ming-Ming Cheng Nankai University The paper 2020/05/08 29 / 46

Page 30: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: honest descriptions

Be honest in your descriptions of your experiments.

Our algorithm to generate music in the style of variousclassical composers worked well for Bach, Scarlatti andMozart, where in each case the test subjects were able toidentify the composer with wiht over 90% success rate (seeTable 1). However, we were less successful at generatingmusic in style of Stravinsky. Users only correctly identifiedthis style in 42% of cases, despite having been given threedifferent pieces by Stravinsky to listen beforehand.

Ming-Ming Cheng Nankai University The paper 2020/05/08 30 / 46

Page 31: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: all methods have limitations

State clearly what the limitations of your method are, andalso give suggestions as to how you might overcome them.

Our new battery technology is designed to operate at roomtemperature. It is unsuitable for outdoors use in cold con-ditions. Our test show that it can only provide 17% of itsmaximum working current at a temperature of 5 °C. It mightbe useful to add an additional temperature control systemto use this battery technology in outdoors.

Ming-Ming Cheng Nankai University The paper 2020/05/08 31 / 46

Page 32: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Experiments: state purpose

Make sure that any conclusions you draw are supportedby the results of your experiments, and that you do notover-generalise.

Our text recognition algorithm works well for English. How-ever, it is unlikely to wrok well for Chinese, which has manymore characters.

Ming-Ming Cheng Nankai University The paper 2020/05/08 32 / 46

Page 33: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Conclusions

At the end of your paper, you should briefly restate yourcontributions and claims to novelty.

Our paper provides novel methods for segmenting a carfrom an image background, and for model-based trackingof cars in video. We show how to use these to estimatethe speed of cars in video, for applications such as lowenforcement and sports broadcasting.

Ming-Ming Cheng Nankai University The paper 2020/05/08 33 / 46

Page 34: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

ConclusionsExplain why your results are significant and useful, andsummarise the experimental evidence, or theoretical proofs,which demonstrate how your ideas are an improvementover ealier work.

The only user interaction required is to indicate when thespeed measurement should be taken. Using a tripod-mounted comsumer-grade camera, an accuracy of 3 milesper hour can be acheived, within 10 seconds of compu-tation. We have experimentally verfied this accuracy bydeploying our system by roadsides and obtaining groundtruth speed by radar. Previous methods have needed user-assistance to find the car in the video, or were less accu-rate.

Ming-Ming Cheng Nankai University The paper 2020/05/08 34 / 46

Page 35: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Conclusions

Also clearly state the limitations of your work.

Our system is generally reliable, but can occasionally failfor grey cars whose colour is too similar to that of road. Im-proved edge detectors may help to overcome this problem.

Ming-Ming Cheng Nankai University The paper 2020/05/08 35 / 46

Page 36: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Acknowledgements

Briefly thank the funders, and e.g. industrial or other part-ners who have provided you with equipment, feedback, orother support. You should also thank other researcherswho have provided you with their datasets, code, speci-mens and so on.

This research was supported by Major Project for NewGeneration of AI under Grant No. 2018AAA0100400and National Science Foundation China under Grant No.61922046. We also wish to thank Prof. Ralph Martin forproviding the example images.

Ming-Ming Cheng Nankai University The paper 2020/05/08 36 / 46

Page 37: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Acknowledgements

Provide citations, links to code/data if possible.If you wish to reproduce the copyrighted works of oth-ers in your paper, it is general not enough to do soand simply mention them in the acknowledgements.You must get the copyright holder’s written permis-sion, altough short extracts of written material may notneed this.

Figure 12. Text recognition results. Left: results producedby Chen’s algorithm (From [16], used by permission ofIEEE. ©IEEE 2009). Right: our results.

Ming-Ming Cheng Nankai University The paper 2020/05/08 37 / 46

Page 38: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Acknowledgements

You may have to pay licence fee to use material.Copyright laws can be complex, and vary from countryto country.Even with permission, you should always state thesource of copyrighted material where you use it.

The copyright needs to be made perfectly clear. Takingpieces of text from other papers (even your other paper)and using them in your own paper is unacceptable, andcounts as plagiarism. Even short passages from other pa-pers should be re-explained in your own words.

Ming-Ming Cheng Nankai University The paper 2020/05/08 38 / 46

Page 39: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Acknowledgements

Fig.: Vectorizing Cartoon Animations, IEEE TVCG 2009.

Ming-Ming Cheng Nankai University The paper 2020/05/08 39 / 46

Page 40: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

References

Bibliography is general list of papers, books, etc.References are items that you have specificallydiscussed and cros-referred to in your paper.Academic papers should contain only references, nota general bibliography.

Ming-Ming Cheng Nankai University The paper 2020/05/08 40 / 46

Page 41: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

ReferencesReferences are used to demonstrate several key pointsand help readers to find out more about the subject.

Show that you are familiar with the field, and itsstate-of-the-art. Only then are your claims of noveltylikely to be plausible.Show how your work is an important over previouswork.

Several previous works have considered three-dimentionalomnispectral cameras [1,2,3], while others have consid-ered four-dimensional hyperspectral cameras [4,5,6,7].Our work is the first to cobmine key elements from bothof these devices, and to show how to construct a four-dimentional omnispectral cameras.

Ming-Ming Cheng Nankai University The paper 2020/05/08 41 / 46

Page 42: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

References

References are used to demonstrate several key pointsand help readers to find out more about the subject.

Tell readers where to find details of any tools andother researchers’ work you have used.Give backbround to your problem.

Our image extrapolation algorithms use combinatorial op-timisation for finding matches between · · · . We use theHungarian algorithm [13] for efficiency.

Du et. al.gives a helpful survey of · · · in [13].

Ming-Ming Cheng Nankai University The paper 2020/05/08 42 / 46

Page 43: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

References format

Adopt the style of citations required by the conference orjournal in which you hope to publish your paper. Somerequire references to be

numbered in alphabetical order.in order of first appearance in the paper.refered by number, e.g. [1].refered by author and date, e.g. [Jones1999].

Chen et. al. \shortcite{Chen09Method} proposed

CONSISTENT and COMPLETE!

Ming-Ming Cheng Nankai University The paper 2020/05/08 43 / 46

Page 44: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Appendices

The main purpose of using appendices is to providesupplementary materials, which is so lengthy that it wouldbreak the reader’s flow of understanding.

An appendix might contain a long proof, where the impor-tant issue is the theorem itself.

An appendix might give a full description of some test data,where an outline will suffice for most readers, but full de-tails are necessary and important for a reader wishing toreplicate the results.

Ming-Ming Cheng Nankai University The paper 2020/05/08 44 / 46

Page 45: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Homework

Submit one page document to disscuss and analysis apaper submission and a PPT presentation. The docu-ment should contain: summary, strengh, weakness, andimprovements recommendations. Use LATEX templatehttps://www.overleaf.com/read/rzdpjzqwkdwb and sub-mit at https://www.wenjuan.com/s/iAJjmen/.

Deadline: 17th May, 2020

Choose one paper from 2 real CVPR submissions.

Deadline: 17th May, 2020

Analysis a real CVPR 2020 presentation (1min).

Ming-Ming Cheng Nankai University The paper 2020/05/08 45 / 46

Page 46: Writting the Paper - 程明明个人主页

Theory Algorithms Experiments Conclusions Acknowledgements References Appendices

Thanks!

Ming-Ming Cheng Nankai University The paper 2020/05/08 46 / 46