Professional Documents
Culture Documents
Ebook Computer Vision Eccv 2020 16Th European Conference Glasgow Uk August 23 28 2020 Proceedings Part Iv Andrea Vedaldi Online PDF All Chapter
Ebook Computer Vision Eccv 2020 16Th European Conference Glasgow Uk August 23 28 2020 Proceedings Part Iv Andrea Vedaldi Online PDF All Chapter
https://ebookmeta.com/product/computer-vision-eccv-2020-16th-
european-conference-glasgow-uk-august-23-28-2020-proceedings-
part-xxvii-andrea-vedaldi/
https://ebookmeta.com/product/computer-vision-eccv-2020-16th-
european-conference-glasgow-uk-august-23-28-2020-proceedings-
part-vi-andrea-vedaldi/
https://ebookmeta.com/product/computer-vision-eccv-2020-16th-
european-conference-glasgow-uk-august-23-28-2020-proceedings-
part-xiii-andrea-vedaldi/
https://ebookmeta.com/product/computer-vision-eccv-2020-16th-
european-conference-glasgow-uk-august-23-28-2020-proceedings-
part-xxix-andrea-vedaldi/
Computer Vision ECCV 2020 16th European Conference
Glasgow UK August 23 28 2020 Proceedings Part VIII
Andrea Vedaldi
https://ebookmeta.com/product/computer-vision-eccv-2020-16th-
european-conference-glasgow-uk-august-23-28-2020-proceedings-
part-viii-andrea-vedaldi/
https://ebookmeta.com/product/computer-vision-eccv-2020-16th-
european-conference-glasgow-uk-august-23-28-2020-proceedings-
part-i-andrea-vedaldi/
https://ebookmeta.com/product/computer-vision-eccv-2020-16th-
european-conference-glasgow-uk-august-23-28-2020-proceedings-
part-xxx-andrea-vedaldi/
https://ebookmeta.com/product/computer-vision-eccv-2020-16th-
european-conference-glasgow-uk-august-23-28-2020-proceedings-
part-xxvi-andrea-vedaldi/
https://ebookmeta.com/product/computer-vision-
eccv-2020-workshops-glasgow-uk-august-23-28-2020-proceedings-
part-ii-lecture-notes-in-computer-science-adrien-bartoli-editor/
Andrea Vedaldi
Horst Bischof
Thomas Brox
Jan-Michael Frahm (Eds.)
LNCS 12349
Computer Vision –
ECCV 2020
16th European Conference
Glasgow, UK, August 23–28, 2020
Proceedings, Part IV
Lecture Notes in Computer Science 12349
Founding Editors
Gerhard Goos
Karlsruhe Institute of Technology, Karlsruhe, Germany
Juris Hartmanis
Cornell University, Ithaca, NY, USA
Computer Vision –
ECCV 2020
16th European Conference
Glasgow, UK, August 23–28, 2020
Proceedings, Part IV
123
Editors
Andrea Vedaldi Horst Bischof
University of Oxford Graz University of Technology
Oxford, UK Graz, Austria
Thomas Brox Jan-Michael Frahm
University of Freiburg University of North Carolina at Chapel Hill
Freiburg im Breisgau, Germany Chapel Hill, NC, USA
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Foreword
Hosting the European Conference on Computer Vision (ECCV 2020) was certainly an
exciting journey. From the 2016 plan to hold it at the Edinburgh International
Conference Centre (hosting 1,800 delegates) to the 2018 plan to hold it at Glasgow’s
Scottish Exhibition Centre (up to 6,000 delegates), we finally ended with moving
online because of the COVID-19 outbreak. While possibly having fewer delegates than
expected because of the online format, ECCV 2020 still had over 3,100 registered
participants.
Although online, the conference delivered most of the activities expected at a
face-to-face conference: peer-reviewed papers, industrial exhibitors, demonstrations,
and messaging between delegates. In addition to the main technical sessions, the
conference included a strong program of satellite events with 16 tutorials and 44
workshops.
Furthermore, the online conference format enabled new conference features. Every
paper had an associated teaser video and a longer full presentation video. Along with
the papers and slides from the videos, all these materials were available the week before
the conference. This allowed delegates to become familiar with the paper content and
be ready for the live interaction with the authors during the conference week. The live
event consisted of brief presentations by the oral and spotlight authors and industrial
sponsors. Question and answer sessions for all papers were timed to occur twice so
delegates from around the world had convenient access to the authors.
As with ECCV 2018, authors’ draft versions of the papers appeared online with
open access, now on both the Computer Vision Foundation (CVF) and the European
Computer Vision Association (ECVA) websites. An archival publication arrangement
was put in place with the cooperation of Springer. SpringerLink hosts the final version
of the papers with further improvements, such as activating reference links and sup-
plementary materials. These two approaches benefit all potential readers: a version
available freely for all researchers, and an authoritative and citable version with
additional benefits for SpringerLink subscribers. We thank Alfred Hofmann and
Aliaksandr Birukou from Springer for helping to negotiate this agreement, which we
expect will continue for future versions of ECCV.
score computed by the Toronto Paper Matching System (TPMS), which is based on the
paper’s full text, the area chair bids for individual papers, load balancing, and conflict
avoidance. Open Review provides the program chairs a convenient web interface to
experiment with different configurations of the matching algorithm. The chosen con-
figuration resulted in about 50% of the assigned papers to be highly ranked by the area
chair bids, and 50% to be ranked in the middle, with very few low bids assigned.
Assignments to reviewers were similar, with two differences. First, there was a
maximum of 7 papers assigned to each reviewer. Second, area chairs recommended up
to seven reviewers per paper, providing another highly-weighed term to the affinity
scores used for matching.
The assignment of papers to area chairs was smooth. However, it was more difficult
to find suitable reviewers for all papers. Having a ratio of 5.6 papers per reviewer with a
maximum load of 7 (due to emergency reviewer commitment), which did not allow for
much wiggle room in order to also satisfy conflict and expertise constraints. We
received some complaints from reviewers who did not feel qualified to review specific
papers and we reassigned them wherever possible. However, the large scale of the
conference, the many constraints, and the fact that a large fraction of such complaints
arrived very late in the review process made this process very difficult and not all
complaints could be addressed.
Reviewers had six weeks to complete their assignments. Possibly due to COVID-19
or the fact that the NeurIPS deadline was moved closer to the review deadline, a record
30% of the reviews were still missing after the deadline. By comparison, ECCV 2018
experienced only 10% missing reviews at this stage of the process. In the subsequent
week, area chairs chased the missing reviews intensely, found replacement reviewers in
their own team, and managed to reach 10% missing reviews. Eventually, we could
provide almost all reviews (more than 99.9%) with a delay of only a couple of days on
the initial schedule by a significant use of emergency reviews. If this trend is confirmed,
it might be a major challenge to run a smooth review process in future editions of
ECCV. The community must reconsider prioritization of the time spent on paper
writing (the number of submissions increased a lot despite COVID-19) and time spent
on paper reviewing (the number of reviews delivered in time decreased a lot pre-
sumably due to COVID-19 or NeurIPS deadline). With this imbalance the peer-review
system that ensures the quality of our top conferences may break soon.
Reviewers submitted their reviews independently. In the reviews, they had the
opportunity to ask questions to the authors to be addressed in the rebuttal. However,
reviewers were told not to request any significant new experiment. Using the Open
Review interface, authors could provide an answer to each individual review, but were
also allowed to cross-reference reviews and responses in their answers. Rather than
PDF files, we allowed the use of formatted text for the rebuttal. The rebuttal and initial
reviews were then made visible to all reviewers and the primary area chair for a given
paper. The area chair encouraged and moderated the reviewer discussion. During the
discussions, reviewers were invited to reach a consensus and possibly adjust their
ratings as a result of the discussion and of the evidence in the rebuttal.
After the discussion period ended, most reviewers entered a final rating and rec-
ommendation, although in many cases this did not differ from their initial recom-
mendation. Based on the updated reviews and discussion, the primary area chair then
Preface ix
made a preliminary decision to accept or reject the paper and wrote a justification for it
(meta-review). Except for cases where the outcome of this process was absolutely clear
(as indicated by the three reviewers and primary area chairs all recommending clear
rejection), the decision was then examined and potentially challenged by a secondary
area chair. This led to further discussion and overturning a small number of preliminary
decisions. Needless to say, there was no in-person area chair meeting, which would
have been impossible due to COVID-19.
Area chairs were invited to observe the consensus of the reviewers whenever
possible and use extreme caution in overturning a clear consensus to accept or reject a
paper. If an area chair still decided to do so, she/he was asked to clearly justify it in the
meta-review and to explicitly obtain the agreement of the secondary area chair. In
practice, very few papers were rejected after being confidently accepted by the
reviewers.
This was the first time Open Review was used as the main platform to run ECCV. In
2018, the program chairs used CMT3 for the user-facing interface and Open Review
internally, for matching and conflict resolution. Since it is clearly preferable to only use
a single platform, this year we switched to using Open Review in full. The experience
was largely positive. The platform is highly-configurable, scalable, and open source.
Being written in Python, it is easy to write scripts to extract data programmatically. The
paper matching and conflict resolution algorithms and interfaces are top-notch, also due
to the excellent author profiles in the platform. Naturally, there were a few kinks along
the way due to the fact that the ECCV Open Review configuration was created from
scratch for this event and it differs in substantial ways from many other Open Review
conferences. However, the Open Review development and support team did a fantastic
job in helping us to get the configuration right and to address issues in a timely manner
as they unavoidably occurred. We cannot thank them enough for the tremendous effort
they put into this project.
Finally, we would like to thank everyone involved in making ECCV 2020 possible
in these very strange and difficult times. This starts with our authors, followed by the
area chairs and reviewers, who ran the review process at an unprecedented scale. The
whole Open Review team (and in particular Melisa Bok, Mohit Unyal, Carlos
Mondragon Chapa, and Celeste Martinez Gomez) worked incredibly hard for the entire
duration of the process. We would also like to thank René Vidal for contributing to the
adoption of Open Review. Our thanks also go to Laurent Charling for TPMS and to the
program chairs of ICML, ICLR, and NeurIPS for cross checking double submissions.
We thank the website chair, Giovanni Farinella, and the CPI team (in particular Ashley
Cook, Miriam Verdon, Nicola McGrane, and Sharon Kerr) for promptly adding
material to the website as needed in the various phases of the process. Finally, we thank
the publication chairs, Albert Ali Salah, Hamdi Dibeklioglu, Metehan Doyran, Henry
Howard-Jenkins, Victor Prisacariu, Siyu Tang, and Gul Varol, who managed to
compile these substantial proceedings in an exceedingly compressed schedule. We
express our thanks to the ECVA team, in particular Kristina Scherbaum for allowing
open access of the proceedings. We thank Alfred Hofmann from Springer who again
x Preface
serve as the publisher. Finally, we thank the other chairs of ECCV 2020, including in
particular the general chairs for very useful feedback with the handling of the program.
General Chairs
Vittorio Ferrari Google Research, Switzerland
Bob Fisher University of Edinburgh, UK
Cordelia Schmid Google and Inria, France
Emanuele Trucco University of Dundee, UK
Program Chairs
Andrea Vedaldi University of Oxford, UK
Horst Bischof Graz University of Technology, Austria
Thomas Brox University of Freiburg, Germany
Jan-Michael Frahm University of North Carolina, USA
Poster Chair
Stephen Mckenna University of Dundee, UK
Technology Chair
Gerardo Aragon Camarasa University of Glasgow, UK
xii Organization
Tutorial Chairs
Carlo Colombo University of Florence, Italy
Sotirios Tsaftaris University of Edinburgh, UK
Publication Chairs
Albert Ali Salah Utrecht University, The Netherlands
Hamdi Dibeklioglu Bilkent University, Turkey
Metehan Doyran Utrecht University, The Netherlands
Henry Howard-Jenkins University of Oxford, UK
Victor Adrian Prisacariu University of Oxford, UK
Siyu Tang ETH Zurich, Switzerland
Gul Varol University of Oxford, UK
Website Chair
Giovanni Maria Farinella University of Catania, Italy
Workshops Chairs
Adrien Bartoli University of Clermont Auvergne, France
Andrea Fusiello University of Udine, Italy
Area Chairs
Lourdes Agapito University College London, UK
Zeynep Akata University of Tübingen, Germany
Karteek Alahari Inria, France
Antonis Argyros University of Crete, Greece
Hossein Azizpour KTH Royal Institute of Technology, Sweden
Joao P. Barreto Universidade de Coimbra, Portugal
Alexander C. Berg University of North Carolina at Chapel Hill, USA
Matthew B. Blaschko KU Leuven, Belgium
Lubomir D. Bourdev WaveOne, Inc., USA
Edmond Boyer Inria, France
Yuri Boykov University of Waterloo, Canada
Gabriel Brostow University College London, UK
Michael S. Brown National University of Singapore, Singapore
Jianfei Cai Monash University, Australia
Barbara Caputo Politecnico di Torino, Italy
Ayan Chakrabarti Washington University, St. Louis, USA
Tat-Jen Cham Nanyang Technological University, Singapore
Manmohan Chandraker University of California, San Diego, USA
Rama Chellappa Johns Hopkins University, USA
Liang-Chieh Chen Google, USA
Organization xiii
Additional Reviewers
1 Introduction
Fig. 1. In this demonstration, the YOLOv2 detector is evaded using a pattern trained
on the COCO dataset with a carefully constructed objective.
must transfer from the digital world into the physical world, from the attacker
model to the victim model, or from models trained on one dataset to another.
In this paper, we study the transferability of attacks on object detectors
across different architectures, classes, and datasets, with the ultimate goal of
generalizing digital attacks to the real-world. Our study has the following goals:
2 Related Work
Attacks on Object Detection and Semantic Segmentation. While there
is a plethora of work on attacking image classifiers [11,23,25], less work has been
Making an Invisibility Cloak 3
done on more complex vision tasks like object detection and semantic segmen-
tation. Metzen et al. demonstrate that nearly imperceptible adversarial pertur-
bations can fool segmentation models to produce incorrect outputs [24]. Arnab
et al. also show that segmentation models are vulnerable to attacks [1], and claim
that adversarial perturbations fail to transfer across network architectures. Xie
et al. introduce Dense Adversary Generation (DAG), a method that produces
incorrect predictions for pixels in segmentation models or proposals in object
detection frameworks [34]. Wei et al. further extend the attack from images to
videos [33]. In contrast to [33,34], which Attacks the classifier stage of object
detectors, Li et al. attack region proposal networks by decreasing the confidence
scores of positive proposals [19]. DPatch causes misclassification of detectors, by
placing a patch that does not overlap with the objects of interest [22]. Li et al.
add imperceptible patches to the background to fool object detectors [18]. Note
that all of these studies focus on digital (as opposed to physical) attacks with a
specific detector, without studying the transferability of attacks. In this paper,
we systematically evaluate a wide range of popular detectors in both the digital
and physical world, and benchmark how attacks transfer in different settings.
Physical Attacks in the Real World. Kurakin et al. took photos of adver-
sarial images with a camera and input them to a pretrained classifier [16]; they
demonstrate that a large fraction of images are misclassified. Eykholt et al. con-
sider physical attacks on stop sign classifiers using images cropped from video
frames [9]. They successfully fool classifiers using both norm bounded perturba-
tions, and also sparse perturbations using carefully placed stickers. Stop signs
attacks on object detectors are considered in [5,8]. Lu et al. showed that the
perturbed sign images from [9] can be reliably recognized by popular detectors
like Faster-RCNN [27] and Yolov2 [26], and showed that detectors are much
more robust to attacks than classifiers. Note that fooling stop sign detectors
differs from fooling person detectors because stop sign perturbations can cover
the whole object, whereas our person attacks leave the face, hands, and legs
uncovered.
Zeng et al. use rendering tools to perform attacks in 3D environments [37].
Sitawarin et al. [30] propose large out-of-distribution perturbations, producing
toxic signs to deceive autonomous vehicles. Athalye et al. introduce expectation
over transformations (EoT) to generate physically robust adversarial samples,
and they produce 3D physical adversarial objects that can attack classifiers in
different conditions. Sharif et al. explore adversarial eyeglass frames that fool
face classifiers [28]. Brown et al. placed adversarial patches [3] on raw images,
forcing classifiers to output incorrect predictions. Komkov et al. generate stick-
ers attached to hats to attack face classifiers [15]. Huang et al. craft attacks
by simulations to cause misclassification of detectors [14]. Thys et al. produce
printed adversarial patches [31] that deceive person detectors instantiated by
Yolov2 [26]. This proof-of-concept study was the first to consider physical attacks
on detectors, although it was restricted to the white-box setting. Furthermore
the authors did not quantify the performance, or address issues like robustness
Another random document with
no related content on Scribd:
line the paragaster as fast as its original covering of choanocytes
retreats into the newly formed chambers.
Fig. 81.—S. setosum. Young Sponge, with one whorl of radial tubes. o, Osculum;
p, pore; sp1, monaxon; sp4, quadriradiate spicule. (After Maas.)
With a canal system precisely similar to that of Sycon, Ute (Fig. 83)
shows an advance in structure in the thickening of the dermal layers
over the distal ends of the chambers. The dermal thickenings above
neighbouring chambers extend laterally and meet; and there results
a sheet of dermal tissue perforated by dermal ostia, which open into
the inhalant canals, and strengthened by stout spicules running
longitudinally. This layer is termed a cortex; it covers the whole
sponge, compacting the radial tubes so that they form, together with
the cortex, a secondary wall to the sponge, which is once more a
simple sac, but with a complex wall. The cortex may be enormously
developed, so as to form more than half the thickness of the wall
(Fig. 84). The chambers taken together are spoken of as the
chamber layer.
Fig. 82.—Sycon raphanus. A, Longitudinal section of young decalcified Sponge
at a stage somewhat later than that shown in Fig. 81. B, Transverse section
of the same through a whorl of tubes. d, Dermal membrane; g, gastral
membrane; H, paragaster; sp4, tetraradiate spicule; T, radial tube. (After
Maas.)
Fig. 83.—Transverse section of the body-wall of Ute, passing longitudinally
through two chambers. a.p, Apopyle; d.o, dermal ostium; fl.ch, flagellated
chamber or radial tube; i.c, inhalant canal; p, prosopyle. (After Dendy.)
Fig. 87.—A spicule from the skeleton framework of Plectroninia, showing the
terminally expanded rays. (After Hinde.)
The sub-family contains only one living genus and a few recently
described fossil forms. Petrostroma schulzei[224] lives in shallow
water near Japan; Plectroninia halli[225] and Bactronella were found
in Eocene beds of Victoria; Porosphaera[226] long known from the
Chalk of England and of the Continent, has recently been shown by
Hinde[226] to be nearly allied to Plectroninia; finally, Plectinia[227] is a
genus erected by Počta for a sponge from Cenomanian beds of
Bohemia. Doederlein, in 1896, expressed his opinion that fossil
representatives of Lithoninae would most surely be discovered. The
fused spicules are equiangular quadriradiates; they are united in
Petrostroma by lateral fusion of the rays, in Plectroninia (Fig. 87) and
Porosphaera by fusion of apposed terminal flat expansions of the
rays, and in some, possibly all, genera a continuous deposit of
calcium carbonate ensheaths the spicular reticulum. Thus they recall
the formation of the skeleton on the one hand of the Lithistida and on
the other of the Dictyonine Hexactinellida (see pp. 202, 211).
"Tuning-forks" may occur in the dermal membrane.
CLASS I. MYXOSPONGIAE
The class Myxospongiae is a purely artificial one, containing widely
divergent forms, which possess a common negative character,
namely, the absence of a skeleton. As a result of this absence they
are all encrusting in habit.
We have said that the Hexactinellids are deep-sea forms; they are
either directly fixed to the bottom or more often moored in the ooze
by long tufts of rooting spicules. In the "glass-rope sponge," the
rooting tuft of long spicules, looking like a bundle of spun glass, is
valued by the Japanese, who export it to us. In Monorhaphis the
rooting tuft is replaced by a single giant spicule,[232] three metres in
length, and described as "of the thickness of a little finger"! Probably
it is as a result of their fixed life in the calm waters of the deep
sea[233] that Hexactinellids contrast with most other sponges by their
symmetry. It should not, however, be forgotten that many of the
Calcarea which inhabit shallow water exhibit almost as perfect a
symmetry.
The skeleton which supports the soft parts is, like them, simple and
constant in its main features. It is secreted by scleroblasts, which lie
in the trabeculae, and is made up of only one kind of spicule and its
modifications. This is the hexactine, a spicule which possesses six
rays disposed along three rectangular axes. Each ray contains an
axial thread, which meets its fellow at the centre of the spicule,
where they together form the axial cross. Modifications of the
hexactine arise either by reduction or branching, by spinulation or
expansion of one or more of the rays. The forms of spicule arising by
reduction are termed pentactines, tetractines, and so on, according
to the number of the remaining rays. Those rays which are
suppressed leave the proximal portion of their axial thread as a
remnant marking their former position (Fig. 94). Octactine spicules
seem to form an exception to the above statements, but Schulze has
shown that they too are but modifications of the hexactine arising by
(1) branching of the rays of a hexactine, followed by (2)
recombination of the secondary rays (Fig. 92).
Fig. 92.—A, discohexaster, in which the four cladi a, a', b, b', c of each ray start
directly from a central nodule. B, disco-octaster, resulting from the
redistribution of the twenty-four cladi of A into eight groups of three. (After
Schulze, from Delage.)
The real or natural cleft in the class lies between those genera
possessing amphidiscs (Figs. 94, 97) among their microscleres, and
all the remainder of the Hexactinellida which bear hexasters (Fig.
96). The former set of genera constitute the sub-class
Amphidiscophora, the latter the Hexasterophora.
Ijima, who has dredged Euplectellids from the waters near Tokyo,
finds that in young specimens oscula are confined to the sieve plate;
parietal gaps are secondary formations. The groundwork of the
skeleton is a lattice similar to that shown in Fig. 100. The chamber-
layer is much folded. Various foreign species of Euplectella afford
interesting examples of association with a Decapod Crustacean,
Spongicola venusta, of which a pair lives in the paragaster of each
specimen. The Crustacean is light pink, the female distinguished by
a green ovary, which can be seen through the transparent tissues. It
is not altogether clear what the prisoner gains, nor what fee, if any,
the host exacts.