How & How Not to Write Scientific

Articles

Students perspective

The Retreat Team

Computational Imaging and Simulation Technologies in Biomedicine (CISTIB)
Infromation and Communications Technology Department
Universitat Pompeu Fabra
Barcelona, Spain
www.cilab.upf.edu

March 17-18, 2008

Publishing is hard
 You are competing with smart people

 CVPR 2007: 4.8% oral, 28.2% overall

 IJCAI 2007: 15.7% oral, 34.7% overall
 ECCV 2006: 4.4% oral, 21.4% overall
 ICCV 2005: 3.7% oral, 19.9% overall

 Watch out for the reviewers!

http://mmlab.ie.cuhk.edu.hk/ 2
The good news

Paper-writing skills can be learned

from good and bad examples

 Reading
 Reviewing
 Writing (and trying to improve)

3
The good news continues ...
Planning
Starting – Dirty your hands
Create a detailed outline of
Implement state-of-the-art
the entire paper

Colleagues
Writing

Checking
Use tools, e.g. dictionaries,
Rewriting
cross-referencing…

4
Today’s take home message

How-to-writing + writing = Acceptance!

= Promotions
= Recognitions

= Jobs =

5
Theme of the talk

“Information is interpreted more easily

and more uniformly if it is placed where
most readers expect to find it”
(Gopen & Swan 90)

6
Abstract

 Typical abstract
 Subject
 Contributions (if any) and methods
 Results

 One fact / idea per sentence

7
Abstract example (Viola & Jones 2001)
This paper describes a visual object detection framework that is capable of
processing images extremely rapidly while achieving high detection rates. There
are three key contributions. The first is the introduction of a new image
representation called the “Integral Image” which allows the features used by our
detector to be computed very quickly. The second is a learning algorithm, based
on AdaBoost, which selects a small number of critical visual features and yields
extremely efficient classifiers [6]. The third contribution is a method for
combining classifiers in a “cascade” which allows background regions of the
image to be quickly discarded while spending more computation on promising
object-like regions. A set of experiments in the domain of face detection are
presented. The system yields face detection performace comparable to the best
previous systems [18, 13, 16, 12, 1]. Implemented on a conventional desktop,
face detection proceeds at 15 frames per second.

8
Abstract example (Viola & Jones 2001)
This paper describes a visual object detection framework that is capable of
processing images extremely rapidly while achieving high detection rates. There
are three key contributions. The first is the introduction of a new image
representation called the “Integral Image” which allows the features used by our
detector to be computed very quickly. The second is a learning algorithm, based
on AdaBoost, which selects a small number of critical visual features and yields
extremely efficient classifiers [6]. The third contribution is a method for
combining classifiers in a “cascade” which allows background regions of the
image to be quickly discarded while spending more computation on promising
object-like regions. A set of experiments in the domain of face detection are
presented. The system yields face detection performace comparable to the best
previous systems [18, 13, 16, 12, 1]. Implemented on a conventional desktop,
face detection proceeds at 15 frames per second.

9
Abstract example (Viola & Jones 2001)
This paper describes a visual object detection framework that is capable of
processing images extremely rapidly while achieving high detection rates. There
are three key contributions. The first is the introduction of a new image
representation called the “Integral Image” which allows the features used by our
detector to be computed very quickly. The second is a learning algorithm, based
on AdaBoost, which selects a small number of critical visual features and yields
extremely efficient classifiers [6]. The third contribution is a method for
combining classifiers in a “cascade” which allows background regions of the
image to be quickly discarded while spending more computation on promising
object-like regions. A set of experiments in the domain of face detection are
presented. The system yields face detection performace comparable to the best
previous systems [18, 13, 16, 12, 1]. Implemented on a conventional desktop,
face detection proceeds at 15 frames per second.

10
Abstract example (Viola & Jones 2001)
This paper describes a visual object detection framework that is capable of
processing images extremely rapidly while achieving high detection rates. There
are three key contributions. The first is the introduction of a new image
representation called the “Integral Image” which allows the features used by our
detector to be computed very quickly. The second is a learning algorithm, based
on AdaBoost, which selects a small number of critical visual features and yields
extremely efficient classifiers [6]. The third contribution is a method for
combining classifiers in a “cascade” which allows background regions of the
image to be quickly discarded while spending more computation on promising
object-like regions. A set of experiments in the domain of face detection are
presented. The system yields face detection performace comparable to the best
previous systems [18, 13, 16, 12, 1]. Implemented on a conventional desktop,
face detection proceeds at 15 frames per second.

11
Abstract example (Viola & Jones 2001)
This paper describes a visual object detection framework that is capable of
processing images extremely rapidly while achieving high detection rates. There
are three key contributions. The first is the introduction of a new image
representation called the “Integral Image” which allows the features used by our
detector to be computed very quickly. The second is a learning algorithm, based
on AdaBoost, which selects a small number of critical visual features and yields
extremely efficient classifiers [6]. The third contribution is a method for
combining classifiers in a “cascade” which allows background regions of the
image to be quickly discarded while spending more computation on promising
object-like regions. A set of experiments in the domain of face detection are
presented. The system yields face detection performace comparable to the best
previous systems [18, 13, 16, 12, 1]. Implemented on a conventional desktop,
face detection proceeds at 15 frames per second.

12
Abstract example (Better one?)
This paper describes a visual object detection framework that is capable of
processing images extremely rapidly while achieving high detection rates. There
are three key contributions. The first is the introduction of a new image
representation called the “Integral Image” which allows the features used by our
detector to be computed very quickly. The second is a learning algorithm, based
on AdaBoost, which selects a small number of critical visual features and yields
extremely efficient classifiers. The third contribution is a method for combining
classifiers in a “cascade” which allows background regions of the image to be
quickly discarded while spending more computation on promising object-like
regions. A set of experiments in the domain of face detection are presented. On
MIT+CMU database, the system achieves a correct detection rate of 0.91 and
false detection rate of 5e-6, which is comparable to the best previous systems.
Implemented on a conventional desktop, face detection proceeds at 15 frames
per second.

13
Abstract example (Better one?)
This paper describes a visual object detection framework that is capable of
processing images extremely rapidly while achieving high detection rates. There
are three key contributions. The first is the introduction of a new image
representation called the “Integral Image” which allows the features used by our
detector to be computed very quickly. The second is a learning algorithm, based
on AdaBoost, which selects a small number of critical visual features and yields
extremely efficient classifiers. The third contribution is a method for combining
classifiers in a “cascade” which allows background regions of the image to be
quickly discarded while spending more computation on promising object-like
regions. A set of experiments in the domain of face detection are presented. On
MIT+CMU database, The system achieves a correct detection rate of 0.91 and
false detection rate of 5e-6, which is comparable to the best previous systems.
Implemented on a computer with the Intel PIII 455 Mhz processor, our face
detector runs at 15 frames per second, approximately 10 times faster than the
alternative approaches.
14
Introduction

 The problem statement

 Why is it interesting and important?
 Why is it hard? (E.g., why do naive approaches fail?)
 Why hasn't it been solved before?

 Review of similar techniques

 Your contributions
 key components

15
Introduction – a more direct approach

 A brief problem statement

 without too many references

 One sentence on what your paper is about

 Review of similar techniques

 Your contributions
 key components

16
Introduction – a more direct approach

In many computer vision problems such as detection, both error

rates and computational complexity reflected by time to
decision, characterize the quality of a given algorithm. We
show that such problems can be formalized in the framework of
sequential decision-making. The optimal strategy in terms of
the shortest average decision time subject to a constraint on
error rates (false positive and false negative rates) is the Wald’s
sequential probability ratio test (SPRT). In the paper, we build
on Wald’s theory and propose an algorithm for two-class
classification problems with near optimal trade-off between
time and error rate.

17
Introduction – a more direct approach
(Sochman & Matas 05)
In many computer vision problems such as detection, both error
rates and computational complexity reflected by time to
decision, characterize the quality of a given algorithm. We
show that such problems can be formalized in the framework of
sequential decision-making. The optimal strategy in terms of
the shortest average decision time subject to a constraint on
error rates (false positive and false negative rates) is the Wald’s
sequential probability ratio test (SPRT). In the paper, we build
on Wald’s theory and propose an algorithm for two-class
classification problems with near optimal trade-off between
time and error rate.

18
Title

 Gives the first impression

 Clearly identifies your work

 Not too long

19
Interesting titles

A novel similarity based quality metric for image fusion

20
Interesting titles

Database of normal human cerebral blood flow,

cerebral blood volume, cerebral oxygen extraction
fraction and cerebral metabolic rate of oxygen
measured by positron emission tomography with 15O-
labelled carbon dioxide or water, carbon monoxide
and oxygen: a multicentre study in Japan

21
A few words about math…

 There is no excuse for math errors.

 Words or multi-letter variables in equations.

 Shorten equations (but don’t skip steps in derivation).

 Use standard notation and terminology as much as possible.

 Helpful LaTeX packages: algorithm2e, pseudocode, amsmath

22
Scalars
Vectors
Elements of a vector
Iterators
Generic indexing
Referring to equations …

Use “(1),” not “Eq. (1)” or “Equation (1),” except at the

beginning of a sentence: “Equation (1) is ...”.

28
Presenting results

 Figures

 Tables
A few words about figures…

 Try to save figures in vector format

 eps for example

 Embed subfigures into one image when possible

30
Comparison of methods (bad)

Surface to surface distance Surface to surface distance

Method 1 vs Ground Truth Method 2 vs Ground Truth 31
Comparison of methods (better)

Surface to surface distance Surface to surface distance

Method 1 vs Ground Truth Method 2 vs Ground Truth 32
Comparison of methods (bad)
Comparison of methods (better)
Colors and resolution (Castro et al. 06)

Original Printed
How to waste space 101 (Gauvrit 05)
How to waste space 101 (Satoh et al. 03)
Incorrect model – we know aneurysms are smooth
(Hassan et al. 03)
Captions

Figure axis labels are often a source of confusion. Try to use

words rather then symbols. As an example write the quantity
“Inductance”, or “Inductance L”, not just. Put units in
parentheses. Do not label axes only with units. In the example,
write “Inductance (mH)”, or “Inductance L (mH)”, not just
“mH”. Do not label axes with the ratio of quantities and units.
For example, write “Temperature (K)”, not “Temperature/K”.

39
 Captions (bad)

Figure 8: Before (a,d) and after (b,e) the registration process.

(c,f) are corresponding maps.
40
 Captions (better)

Figure 9: Results of the registration to a reference applied to three

histograms.
(a,d,g) Pairs of histograms before the registration process.
(b,e,h) The pairs of histograms after the registration process
(c,f,i) The corresponding deformation maps. 41
Table Design (bad)

 Messy
 Not so much information for such large space
Table Design (better)

 Simpler
 Cleaner
 More information, less space
How to waste space 101 (McKinney 07)
Language

Interesting read:
“The science of scientific writing” (Gopen & Swan 90)

 Science is often hard to read.

 Usually difficulties in understanding are NOT born out of

complexity of scientific concepts

 Clarity in communication without oversimplifying scientific

issues

47
Language

 Write without grammatical errors

 Correct use of tense

 Be polite

48
Verb Tenses

 Past Tense

 Description of the materials / methods used in the experiment

 Attributing work to another author

 Example

 Correct: “... regions were cropped manually ....”

 Correct: “In [17], local curvatures were used as features ...”
Verb Tenses

 Present Tense

 Background / accepted information

 Description of figures
 Discussion of results and drawing conclusions

 Example

 Correct: “... feature is composed of two regions ...”

 Correct: “…is shown in Fig. 1.”
 Correct: “Experimental results show that our method yields ...”
Active vs. Passive Voice

Imagine you are writing an essay on energy crisis

Passive:

In the 1980s, the interest in the direct conversion of methane to

methanol was renewed by the “energy crisis” and the demand for the
efficient utilization of abundant natural gas reserves.

Active:

The “energy crisis” and the demand for the efficient utilization of
abundant natural gas reserves renewed the interest…
Active vs. Passive Voice

 Passive voice is sometimes appropriate, but should not be

used often.

 If the sentence can be rewritten with the subject performing

the action, do it!
Sentence Length

 Judge each sentence individually.

 Too short – combine two adjacent sentences

 Too long – look for conjunctions and split the sentence

Sentence Length (cont.)
 Particularly bad example:

The result of measurement of the apparent activation

energy predicted that doping of a minor amount of the
CNTs into the CuiZnjAlk did not alter the reaction pathway
of hydrogenation of CO/CO2 to form methanol, however,
led to a considerable increase in the active Cu surface area
of the catalyst and pronounced enhance of the stationary-
state concentration of active hydrogen-adspecies on the
surface of the functioning catalyst, as well as 10~20
degrees dropping down of optimum operation temperature
required for methanol synthesis; all these would contribute
considerably to an increase in reaction activity of
methanol synthesis.
Sentence Length (cont.)
 Improved

Measurements of the apparent activation energy predict that

doping of a minor amount of CNTs into the CuiZnjAlk does
not alter the hydrogenation of CO/CO2 reaction pathway to
form methanol. However, this did lead to a considerable increase
in the active Cu surface area of the catalyst and a pronounced
enhancement of the stationary-state concentration of active
hydrogen-adspecies on the surface of the functioning catalyst.
In addition, the optimum operating temperature required for
methanol synthesis drops 10~20 degrees. All these factors
greatly contribute to the increase in the methanol synthesis
reaction activity.

55
“And” and “But”

Example:

Supports with a high surface area produced high co-dispersions

and highly active FT catalysts. And the nature of the higher
hydrocarbons…

Huang and Ho [24] pioneered the investigation of coupling

modes between shaft torsion and blade bending in a shaft–disk–
blades unit via the receptance method. But the disk was
reasonably assumed rigid due to zero stagger angle.

57
Other mistakes (cont.)

“It is obvious that…” You are using

“It can be seen that…” too many words!
“It is evident that…”

It was found that only manganese and tin oxides were stable…

58
Other mistakes
An example of an bad paper:

An analysis of the frobnicatable foo lter. In this paper we present a

performance analysis of our previous paper [1], and show it to be inferior
to all previously known methods. Why the previous paper was accepted
without this analysis is beyond me.

An example of an excellent paper:

An analysis of the frobnicatable foo lter. In this paper we present a

performance analysis of the paper of Smith [1], and show it to be inferior
to all previously known methods. Why the previous paper was accepted
without this analysis is beyond me.
 Language (Bland & Altman 03)

Estimates are usually made with some sampling error, and limits of
agreement are no exception. We showed how to estimate CIs for the
limits of agreement. Another regret is that these Cis are seldom
quoted. For the data of Cicero et al.12 the mean difference was 0.2
mm with SD 3.0 mm, giving 95% limits of agreements −5.8 to +6.1
mm. There were 231 cases. The standard error of the limits is
approximately3s2/n. This gives3 × 3.02/231 = 0.34. The 95% CI for
the limits of agreement is given by ±1.96 standard errors = 0.67, so for
the lower limit the CI is −6.5 to −5.1 and for the upper limit the 95%
CI is +5.4 to +6.8. Not so hard, really!

60
 Language (Bland & Altman 03)

Estimates are usually made with some sampling error, and limits of
agreement are no exception. We showed how to estimate CIs for the
limits of agreement. Another regret is that these Cis are seldom
quoted. For the data of Cicero et al.12 the mean difference was 0.2
mm with SD 3.0 mm, giving 95% limits of agreements −5.8 to +6.1
mm. There were 231 cases. The standard error of the limits is
approximately3s2/n. This gives3 × 3.02/231 = 0.34. The 95% CI for
the limits of agreement is given by ±1.96 standard errors = 0.67, so for
the lower limit the CI is −6.5 to −5.1 and for the upper limit the 95%
CI is +5.4 to +6.8. Not so hard, really!

61
Language (Bland & Altman 03)

The results section of the paper contains no limits of agreement, but

rather correlation and rank correlation coefficients with P-values!

62
Conclusions

 Typical conclusion section

 Summary and applications
 Main results
 Future work / Open issues

 One major finding per paragraph

Conclusions (Sochman & Matas 05)
In this paper, the two-class classification problems with a
decision quality and time trade-off are formulated in the
framework of the sequential decision-making. We adopted
the optimal SPRT test and enlarged its applicability to Summary
problems with dependent measurements.

In the proposed WaldBoost algorithm, the measurements

are selected and ordered by the AdaBoost algorithm. The
joint probability density function is approximated by the
class-conditional response of the sequence of strong
classifiers. To reduce the effect of inaccurate Method
approximation in early cycles of training, a conservative
method using Parzen windows with a kernel width set
according to the oversmoothing rule was used.

The proposed algorithm was tested on the face detection

problem. On a standard dataset, the results are superior to
the state-of-the-art methods in average evaluation time
andcomparable in detection rates. In the face detection
context, the WaldBoost algorithm can be also viewed as a
Results
theoretically justifiable ”boosted cascade of classifiers”
proposed by Viola and Jones [8].
Conclusions (Sochman & Matas 05)
In this paper, the two-class classification problems with a
decision quality and time trade-off are formulated in the
framework of the sequential decision-making. We adopted
the optimal SPRT test and enlarged its applicability to Summary
problems with dependent measurements.

In the proposed WaldBoost algorithm, the measurements

are selected and ordered by the AdaBoost algorithm. The
joint probability density function is approximated by the
class-conditional response of the sequence of strong
classifiers. To reduce the effect of inaccurate Method
approximation in early cycles of training, a conservative
method using Parzen windows with a kernel width set
according to the oversmoothing rule was used.

The proposed algorithm was tested on the face detection

problem. On a standard dataset, the results are superior to
the state-of-the-art methods in average evaluation time and
comparable in detection rates. In the face detection context,
the WaldBoost algorithm can be also viewed as a
Results
theoretically justifiable ”boosted cascade of classifiers”
proposed by Viola and Jones [8].
References

 Never cite a paper for which you haven’t read at least the
relevant part.

 Don’t cite textbooks – they may be difficult for reader to access,

information may be buried.

 Citing yourself or your group extensively.

 References should be helpful to the reader, not of historical

interest (unless you’re writing a review)

67
References
 We all know how to use variables…  BibTeX provides the @string entry
 Like @inproceedings, @article,
@techreport, @thesis

 Journal=“IEEE TMI” @string(IEEE_J_MI = “IEEE Trans.

 Journal=“IEEE Trans. Med. Imag.” Med. Imag.”)
 Journal=“IEEE Transactions on
Medical  Journal=IEEE_J_MI
Imaging”

68
Using table of contents

 Is there a smooth flow of thoughts?

 Simple with LaTeX: \

tableofcontents
Using table of contents – a better one
Cata’s cure for writer's Block

 Make a first draft even

if it is empty

 Show it to colleagues
and get feedback

71
Today’s take home message

Learn from others!

72
Today’s take home message

 Readers have relatively fixed expectations about where to find

information.

 Information is interpreted more easily if it is placed in the

expected places.

 Awareness of such locations = better control of emphasis a

reader will give to the various
pieces of information.
 Not hard, really!
Happy paper writing!

Thank You!

74
More Materials
 How To Write A Scientific Paper
http://www.improb.com/airchives/paperair/volume2/v2i5/ho
wto.htm

 How To Write Papers, by Charles Ling

 Grammer, Punctuation, and Capitalization

http://www.sti.nasa.gov/publish/sp7084.pdf
http://www.ifc.dicp.ac.cn/scipress/common.ppt

 A Guide for Writing Research Papers Based on Modern

Language Association (MLA) Documentation
http://webster.commnet.edu/mla/index.shtml
75