O slideshow foi denunciado.
Utilizamos seu perfil e dados de atividades no LinkedIn para personalizar e exibir anúncios mais relevantes. Altere suas preferências de anúncios quando desejar.

How to write a clear paper

General tips on organizing paper contents, preparing figures and tables, and writing.

  • Entre para ver os comentários

How to write a clear paper

  1. 1. How to Write a Clear Paper Jia-Bin Huang Virginia Tech www.jiabinhuang.com
  2. 2. “Bad thinking never produces good writing” Good writing promotes good thinking … Leslie Lamport Slide credit: Dimitri Bertsekas
  3. 3. Writing a paper sharpens your mental focus like nothing else. “Like Japanese bonsai art or building a ship in a bottle, paper writing forces you to obsess over every meticulous detail -- word choice, overall tone, readability of graphs -- and of course more mundane details like font size and line spacing. This microscopic attention to every aspect of your work brings out a wonderful, if somewhat exhausting, intellectual rapture. I have never thought so clearly about a piece of research than when I'm in the throes of putting together a paper against a deadline. http://matt-welsh.blogspot.com/2012/07/in-defense-of-scientific-paper.html
  4. 4. This talk • Paper organization • Figures and tables • Notes on writing
  5. 5. How to write a paper •Don Geman’s advice: A paper has four parts 1. Title 2. Abstract 3. Introduction 4. Rest of the paper Spend equal time on all four of these! Slide credit: Jitendra Malik
  6. 6. On the title •Should capture what is special about the paper •Title -> content of the paper. •Recalling the title should help recall the paper •If every other paper in the field could use the same title, it is a bad title •Spend some time inventing terms-memorable and specific are best Slide credit: Jitendra Malik
  7. 7. Paper Organization
  8. 8. Paper organization •Introduction •Related Work •Method •Experimental Results •Conclusions
  9. 9. Introduction section • “The most important section of a paper. For me, once I have finished reading the introduction, I have formed an opinion of whether to accept or reject the paper” – Jitendra Malik • “You must make your paper easy to read. You've got to make it easy for anyone to tell what your paper is about, what problem it solves, why the problem is interesting, what is really new in your paper (and what isn't), why it's so neat. And you must do it up front. In other words, you must write a dynamite introduction.” - Jim Kajiya
  10. 10. Introduction template • Problem, context, importance • Why we should care about this problem? • What did you do (input/output)? Figure-Teaser • Existing approaches and why they are not satisfactory? • If they were satisfactory, you don’t need to do this work. • Figure-Motivation and/or Figure-ToyExample • Our work • Use “In this paper, …” to set a clean separation between existing work and your work. • Our contributions • Concise statement of the novelty of this paper. • Forward reference to sections.
  11. 11. Figure - Teaser • Input and Output [Isola et al 2017] [Darabi et al. 2012] [Huang et al 2016] [Zhang et al 2016]
  12. 12. Figure – Motivation • Examples that highlight the Key Idea of the paper [Parikh and Grauman 2011] [Huang et al. 2015] [Torralba and Efros 2011]
  13. 13. Figure – ToyExample [Simoncelli et al. 1992] [Freeman and Adelson 1991] Slide credit: Bill Freeman
  14. 14. Related Work section • Group existing work into a set of topics • Compare, organize, Table-RelatedWork • Do not just “list” the papers, “relate” them • Make distinction between competing and related (orthogonal) previous work.
  15. 15. Two paragraphs for each topic 1. Topic • What the topic problem is about? Input/output • What is the research trajectory? • Cite relevant papers within your storyline 2. Relationship with this work • Similar • We also adopt … • We build upon … • We share similar … • Different • Unlike/In contrast to … • Our method differs in … • On the other hand, our main focus …
  16. 16. Table – Related Work • Provide conceptual differences to related work [Zhang et al 2017] [Lai et al 2016]
  17. 17. Method section • Use more informative section title than “method”, “algorithm” • Start with an overview subsection. • Define math notations • Provide forward references to subsequent sections • Describe Figure-Overview • Organize the contents into subsections and paragraphs
  18. 18. Figure – Overview • Visualize the algorithm • Provide forward references to equations and sections [Girshick 2015][Xue et al. 2015] [Wadhwa et al. 2013] [Huang et al. 2016]
  19. 19. Results section • Experimental setup • Datasets • Metrics • Compared methods • Implementation details • Training details • Hyper-parameter (data: train/validation/test) • Experiments on dataset A, B, C • Ablation study • Answer questions • Failure modes
  20. 20. Experimental setup • Datasets • Be specific, e.g., which KITTI? which data split? • Number of training/validation/testing images • Metrics • Provide definitions if they are not standard • Compared methods • Group them into meaningful sets, e.g., methods use A, different feature backbone, supervised/unsupervised methods
  21. 21. Implementation details • Training details • Optimizer? • Learning rate? Weight decay? • Training epoch/iterations? Batch size? • Training time? On what machine? • Inference details • Pre-processing data? Resizing? • Hyper-parameters • How did you determine the hyper-parameter values? • Training/validation/testing • Random sampling/grid search • Cross-validation
  22. 22. Results • Visual results • Zoom-in for showing important regions • Examples matter! • Quantitative results • Factorize results according to training data, feature backbone, level of supervision LapSRN [Lai et al. 2018] DF-Net [Zou et al. 2018]
  23. 23. Ablation study • Goal: answer interesting questions with Table-Ablation • One table/figure per question • Does XXX help? • Effect of the number of training data • Effect of parameter X
  24. 24. Table – Ablation Study Non-local neural network [Wang et al. 2018]
  25. 25. Failure modes • Reveal limitations of the method • Speed? memory? model size? • What kind of inputs will break the algorithm? • Reviewer: • “The authors are fully transparent by evoking themselves some limitations in a dedicated paragraph.”
  26. 26. Conclusions/Future work? • Conclusions, or what this opens up, or how this can change how we approach computer vision problems? • “I can’t stand “future work” sections. It’s hard to think of a weaker way to end a paper. “ • “Here’s a list all the ideas we wanted to do but couldn’t get to work in time for the conference submission deadline. We didn’t do any of the following things: (1)...” • (You get no “partial credit” from reviewers and readers for neat things you wanted to do, but didn’t.) • “Here’s a list of good ideas that you should now go and do before we get a chance.” Slide credit: Bill Freeman
  27. 27. Figures and Tables
  28. 28. Figures, tables • If you pull out all the figures and tables and put them into a slide deck you should have a good talk ready • The best way to write a paper is to first give a talk on it. Slide credit: Jitendra Malik
  29. 29. Figures • File format • DO NOT use JPEG images (to avoid compression artifacts). Use PNG or PDF • Resolution • DO NOT use low-resolution images • Position • Put the figures to the top of each page begin{figure}[t] • Caption • The image caption should be self-contained • Highlight the topic of the figure with bold font textbf [Faktor and Irani 2014]
  30. 30. Multiple Images • Use subfigure or minipage. DO NOT use tabular. • Never manually define the physical size of the image • includegraphics[width=5cm]{IMAGE.png} -> Bad • includegraphics[width=0.5linewidth]{IMAGE.png} -> Good • setlength{figwidth}{0.5linewidth} -> Best begin{minipage}{figwidth} includegraphics[width=linewidth]{IMAGE.png} end{minipage}
  31. 31. Multiple Images • Put sub-captions directly under subfigures, do not put them in the caption • All the legends, axis, labels must be clearly visible • Make use of color and textures to code information (a) (b) PatchMatch propagation Flow-guided propagation [Huang et al. 2016]
  32. 32. Spacing between Images begin{figure}[t] % Maximum length includegraphics[width=0.3linewidth]{A.png} hfill includegraphics[width=0.3linewidth]{A.png} % Equal length hspace*{fill} includegraphics[width=0.3linewidth]{B.png} hfill includegraphics[width=0.3linewidth]{B.png} hspace*{fill} % Fixed length centering includegraphics[width=0.3linewidth]{C.png} hspace{1em} includegraphics[width=0.3linewidth]{C.png} end{figure}
  33. 33. Image, video, and dataset names • Use textsc{Name} to separate images, videos, dataset names from the main texts. [Kopf 2016]
  34. 34. Multiple Images • How do I align images with different sizes? • Solve a simple algebra problem • Suppose we know the image on the left has aspect ratio = H/W = c • What’s 𝑥 ? ImA ImB ImB 𝑥 𝑐𝑥 1 − 𝑥 1 − 𝑥 1 − 𝑥 𝑐𝑥 = 2 1 − 𝑥 2 + 𝑐 𝑥 = 2 𝑥 = 2/(2 + c) setlength{figa}{0.612textwidth} setlength{figb}{0.388textwidth} begin{minipage}{figa} includegraphics[width=linewidth]{ImA.png} end{minipage} begin{minipage}{figb} includegraphics[width=linewidth]{ImB.png} includegraphics[width=linewidth]{ImB.png} end{minipage}
  35. 35. Tables – Making nice tables • Which one looks better? Source: Small Guide to Making Nice Tables by Markus Püschel (ETH Zürich) Recommended by David J. Crandall
  36. 36. Algorithms • See the documentation of algorithm2e • Provide the main steps of the algorithm • Use consistent annotations • Use references to sections and equations to connect the main texts with the algorithm [Huang et al. 2016]
  37. 37. Equations • Use begin{equation}…end{equation} environment. • Use begin{algin} … end{align} if you have multiple lines of equations • Label every equation label{eqn:Eqn-Name} • For in-text math symbols, use $$, e.g. Let $x$ be … • Define every notation • For texts that are not part of the equation, use mathrm, e.g. $x_mathrm{color}$
  38. 38. Equations • Number all equations • Easy to refer to them • Equations are grammatical parts of the sentences • Never forget a period after an equation • Never create a dangling displayed equation • Negative numbers • “-” indicate the dash. Use $-1$ to represent minus one • Angle braskets • Use langle and rangle, instead of the comparison operators < and > • Big parentheses • Use left and right for automatic resizing round (), square [], and angled langlerangle brackets as well as vertical bars vert and Vert Source: https://www.cs.dartmouth.edu/~wjarosz/writing.html
  39. 39. Math writing • Mathspeak should be “readable” • BAD: Let k>0 be an integer. • GOOD: Let k be a positive integer or Consider an integer k>0. • BAD: Let x ∈ R^n be a vector. • GOOD: Let x be a vector in R^n or Consider a vector x ∈ R^n • Use simple notation • x(m,n) vs x_{m,n} • Consistent format • Bad example: Proposition 1: If A and B hold, then C and D hold. Proposition 2: C’ and D’ hold, assuming that A’ and B’ are true. • Good example: Proposition 1: If A and B hold, then C and D hold. Proposition 2: If A’ and B’ hold, then C’ and D’ hold. Source: http://www.robots.ox.ac.uk/~phst/Style/Ten_Rules.pdf
  40. 40. Math writing • Keep the reader informed • Where you are and where you are going • Start each segment with a short introduction and perhaps a road map • Announce your intentions/results, e.g., “It turns out that so-and-so is true. To see this, note …” • Tell them what you told them • Refer to equations/results/assumptions by content/name (in addition to number) • e.g., Bellman’s equation, weak duality theorem, etc Source: http://www.robots.ox.ac.uk/~phst/Style/Ten_Rules.pdf
  41. 41. Notes on Writing
  42. 42. Notes on writing • Be concise • Be consistent • Be specific • Be direct • Common mistakes
  43. 43. Be concise BAD GOOD the question as to whether whether (the question whether) there is no doubt but that no doubt (doubtless) used for fuel purposes used for fuel he is a man who he in a hasty manner hastily this is a subject which this subject His story is a strange one. His story is strange. Element of Style, Stunk and White
  44. 44. No “the fact that” BAD GOOD owing to the fact that since (because) in spite of the fact that though (although) call your attention to the fact that remind you (notify you) I was unaware of the fact that I was unaware that (did not know) the fact that he had not succeeded his failure the fact that I had arrived my arrival Element of Style, Stunk and White
  45. 45. Be consistent • BAD: Formerly, science was taught by the textbook method, while now the laboratory method is employed. • GOOD: Formerly, science was taught by the textbook method; now it is taught by the laboratory method. • Correlative expressions (both, and; not, but; not only, but also; either, or; first, second, third; and the like) should be followed by the same grammatical construction. BAD GOOD It was both a long ceremony and very tedious. The ceremony was both long and tedious. A time not for words, but action A time not for words, but for action Either you must grant his request or incur his ill will. You must either grant his request or incur his ill will. Element of Style, Stunk and White
  46. 46. Be specific and direct • Search all the pronoun “which, it, they, that” in the paper • Remove ambiguity by replacing them with the actual noun • Direct: use active tense
  47. 47. Wording BAD GOOD Recent years In recent years Test the method Verify the method Verify the correctness of Validate/demonstrate the The scheme works as The scheme perfoms as Follow [1], we … Following Huang et al. [1], we … Very often Oftentimes / Usually In summary Overall Artificial images Synthetic images As mentioned earlier As stated in Section X Popular algorithms Existing algorithms Slide credit: Ming-Hsuan Yang
  48. 48. Wording BAD GOOD Build layers Construct layers Outperforms Performs favorably against Besides Furthermore / In addition Problems are still not clear Problems remain unclear Many efforts Much efforts Performs the best Perform best The reviewer has comments on The reviewer doubts … We search similar regions We search for similar regions Slide credit: Ming-Hsuan Yang
  49. 49. Slide credit: Ming-Hsuan Yang
  50. 50. Slide credit: Ming-Hsuan Yang
  51. 51. Slide credit: Ming-Hsuan Yang
  52. 52. Slide credit: Ming-Hsuan Yang
  53. 53. Slide credit: Ming-Hsuan Yang
  54. 54. Slide credit: Ming-Hsuan Yang
  55. 55. Slide credit: Ming-Hsuan Yang
  56. 56. If you don’t remember anything else • Writing is an iterative process • Paper organization • Introduction, Related Work, Method, Result, Conclusion • Figures and tables • Give a talk first • Polish them • Notes on writing • Attention to details • Be concise, direct, specific
  57. 57. Resources • Awesome computer vision – writing by Jia-Bin Huang (Virginia Tech) • A quick guide to LaTeX by Dave Richeson (Dickinson College) • Common mistakes in technical writing by Wojciech Jarosz (Dartmouth College) • SIGGRAPH paper template by Li-Yi Wei (University of Hong Kong) • Notes on writing by Fredo Durand (MIT) • How to write a good CVPR submission by Bill Freeman (MIT) • How to write a great research paper by Simon Peyton Jones (MSR) • How to write papers so people can read them by Derek Dreyer (MPI)