Paper Gestalt
Carven von Bearnensquash
Department of Computer Science
University of Phoenix
bearensquash@live.com
Abstract
Peer reviews of conference paper submissions is an in-
tegral part of the research cycle, though it has unknown
origins. For the computer vision community, this process
has become significantly more difficult in recent years due
to the volume of submissions. For example, the number of
submissions to the CVPR conference has tripled in the last
ten years. For this reason, the community has been forced to
reach out to a less than ideal pool of reviewers, which un-
fortunately includes uninformed junior graduate students,
disgruntled senior graduate students, and tenured faculty.
In this work we take the simple intuition that the quality
of a paper can be estimated by merely glancing through
the general layout, and use this intuition to build a system
that employs basic computer vision techniques to predict if
the paper should be accepted or rejected. This system can
then be used as a first cascade layer during the review pro-
cess. Our results show that while rejecting 15% of “good
papers”, we can cut down the number of “bad papers” by
more than 50%, saving valuable time of reviewers. Finally,
we fed this very paper into our system and are happy to
report that it received a posterior probability of 88.4% of
being “good”.
1. Introduction
Peer reviews of conference paper submissions is an in-
tegral part of the research cycle, though it has unknown
origins. For the computer vision community, this process
has become significantly more difficult in recent years due
to the volume of submissions. For example, the number
of submission to the CVPR conference has tripled in the
last ten years
1
(see Fig. 1). For this reason, the commu-
nity has been forced to reach out to a less than ideal pool
of reviewers, which unfortunately includes uninformed ju-
nior graduate students, disgruntled senior graduate students,
1
http://www.adaptivebox.net/CILib/CICON_stat.
html.
Figure 1. Paper submission trends. The number of submitted
papers to CVPR, and other top tier computer vision conferences,
is growing at an alarming rate. In this paper we propose an au-
tomated method of rejected sub-par papers, thereby reducing the
burden on reviewers.
and tenured faculty. Although many excellent research pa-
pers have been published in the area of computer vision
[3, 14, 15, 11, 19, 8, 7, 16, 1, 18, 17, 5, 2, 20, 22, 13], many
good papers are rejected and many bad papers are accepted
due to the imperfect review process.
In this work we take the simple intuition that the quality
of a paper can be estimated by merely glancing through the
general layout, and use this intuition to build a system that
emplys basic computer vision techniques to predict if the
paper should be accepted or rejected. We call the set of
visual features that have discriminative power in this task
the “paper gestalt”. To build our system, we use powerful
statistical learning techniques [4].
The rest of the paper is laid out as follows. In Section 2
we discuss some of the related work (though our work is so
unique that there is little to discuss). In Section 3 we review
our particular solution to this difficult problem. In Section
4 we present our thorough experimental results. Finally, in
1