Professional Documents
Culture Documents
Computer Vision ECCV 2018 15th European Conference Munich Germany September 8 14 2018 Proceedings Part I Vittorio Ferrari
Computer Vision ECCV 2018 15th European Conference Munich Germany September 8 14 2018 Proceedings Part I Vittorio Ferrari
https://textbookfull.com/product/computer-vision-eccv-2018-15th-
european-conference-munich-germany-
september-8-14-2018-proceedings-part-iii-vittorio-ferrari/
https://textbookfull.com/product/computer-vision-eccv-2018-15th-
european-conference-munich-germany-
september-8-14-2018-proceedings-part-iv-vittorio-ferrari/
https://textbookfull.com/product/computer-vision-eccv-2018-15th-
european-conference-munich-germany-
september-8-14-2018-proceedings-part-vii-vittorio-ferrari/
https://textbookfull.com/product/computer-vision-eccv-2018-15th-
european-conference-munich-germany-
september-8-14-2018-proceedings-part-x-vittorio-ferrari/
Computer Vision ECCV 2018 15th European Conference
Munich Germany September 8 14 2018 Proceedings Part
XIII Vittorio Ferrari
https://textbookfull.com/product/computer-vision-eccv-2018-15th-
european-conference-munich-germany-
september-8-14-2018-proceedings-part-xiii-vittorio-ferrari/
https://textbookfull.com/product/computer-vision-eccv-2018-15th-
european-conference-munich-germany-
september-8-14-2018-proceedings-part-xiv-vittorio-ferrari/
https://textbookfull.com/product/computer-vision-
eccv-2018-workshops-munich-germany-
september-8-14-2018-proceedings-part-ii-laura-leal-taixe/
https://textbookfull.com/product/computer-vision-
eccv-2018-workshops-munich-germany-
september-8-14-2018-proceedings-part-v-laura-leal-taixe/
https://textbookfull.com/product/computer-vision-eccv-2016-14th-
european-conference-amsterdam-the-netherlands-
october-11-14-2016-proceedings-part-i-1st-edition-bastian-leibe/
Vittorio Ferrari · Martial Hebert
Cristian Sminchisescu
Yair Weiss (Eds.)
LNCS 11205
Computer Vision –
ECCV 2018
15th European Conference
Munich, Germany, September 8–14, 2018
Proceedings, Part I
123
Lecture Notes in Computer Science 11205
Commenced Publication in 1973
Founding and Former Series Editors:
Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen
Editorial Board
David Hutchison
Lancaster University, Lancaster, UK
Takeo Kanade
Carnegie Mellon University, Pittsburgh, PA, USA
Josef Kittler
University of Surrey, Guildford, UK
Jon M. Kleinberg
Cornell University, Ithaca, NY, USA
Friedemann Mattern
ETH Zurich, Zurich, Switzerland
John C. Mitchell
Stanford University, Stanford, CA, USA
Moni Naor
Weizmann Institute of Science, Rehovot, Israel
C. Pandu Rangan
Indian Institute of Technology Madras, Chennai, India
Bernhard Steffen
TU Dortmund University, Dortmund, Germany
Demetri Terzopoulos
University of California, Los Angeles, CA, USA
Doug Tygar
University of California, Berkeley, CA, USA
Gerhard Weikum
Max Planck Institute for Informatics, Saarbrücken, Germany
More information about this series at http://www.springer.com/series/7412
Vittorio Ferrari Martial Hebert
•
Computer Vision –
ECCV 2018
15th European Conference
Munich, Germany, September 8–14, 2018
Proceedings, Part I
123
Editors
Vittorio Ferrari Cristian Sminchisescu
Google Research Google Research
Zurich Zurich
Switzerland Switzerland
Martial Hebert Yair Weiss
Carnegie Mellon University Hebrew University of Jerusalem
Pittsburgh, PA Jerusalem
USA Israel
LNCS Sublibrary: SL6 – Image Processing, Computer Vision, Pattern Recognition, and Graphics
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Foreword
It was our great pleasure to host the European Conference on Computer Vision 2018 in
Munich, Germany. This constituted by far the largest ECCV event ever. With close to
2,900 registered participants and another 600 on the waiting list one month before the
conference, participation more than doubled since the last ECCV in Amsterdam. We
believe that this is due to a dramatic growth of the computer vision community
combined with the popularity of Munich as a major European hub of culture, science,
and industry. The conference took place in the heart of Munich in the concert hall
Gasteig with workshops and tutorials held at the downtown campus of the Technical
University of Munich.
One of the major innovations for ECCV 2018 was the free perpetual availability of
all conference and workshop papers, which is often referred to as open access. We note
that this is not precisely the same use of the term as in the Budapest declaration. Since
2013, CVPR and ICCV have had their papers hosted by the Computer Vision Foun-
dation (CVF), in parallel with the IEEE Xplore version. This has proved highly ben-
eficial to the computer vision community.
We are delighted to announce that for ECCV 2018 a very similar arrangement was
put in place with the cooperation of Springer. In particular, the author’s final version
will be freely available in perpetuity on a CVF page, while SpringerLink will continue
to host a version with further improvements, such as activating reference links and
including video. We believe that this will give readers the best of both worlds;
researchers who are focused on the technical content will have a freely available
version in an easily accessible place, while subscribers to SpringerLink will continue to
have the additional benefits that this provides. We thank Alfred Hofmann from
Springer for helping to negotiate this agreement, which we expect will continue for
future versions of ECCV.
refining conflict definitions and in generating quality matches. The only glitch is that,
once the matches were generated, a small percentage of papers were unassigned
because of discrepancies between the OpenReview conflicts and the conflicts entered in
CMT3. We manually assigned these papers. This glitch is revealing of the challenge of
using multiple systems at once (CMT3 and OpenReview in this case), which needs to
be addressed in future.
After assignment of papers to ACs, the ACs suggested seven reviewers per paper
from the Program Committee pool. The selection and rank ordering were facilitated by
the TPMS affinity scores visible to the ACs for each paper/reviewer pair. The final
assignment of papers to reviewers was generated again through OpenReview in order
to account for refined conflict definitions. This required new features in the OpenRe-
view matching system to accommodate the ECCV workflow, in particular to incor-
porate selection ranking, and maximum reviewer load. Very few papers received fewer
than three reviewers after matching and were handled through manual assignment.
Reviewers were then asked to comment on the merit of each paper and to make an
initial recommendation ranging from definitely reject to definitely accept, including a
borderline rating. The reviewers were also asked to suggest explicit questions they
wanted to see answered in the authors’ rebuttal. The initial review period was five
weeks. Because of the delay in getting all the reviews in, we had to delay the final
release of the reviews by four days. However, because of the slack included at the tail
end of the schedule, we were able to maintain the decision target date with sufficient
time for all the phases. We reassigned over 100 reviews from 40 reviewers during the
review period. Unfortunately, the main reason for these reassignments was reviewers
declining to review, after having accepted to do so. Other reasons included technical
relevance and occasional unidentified conflicts. We express our thanks to the emer-
gency reviewers who generously accepted to perform these reviews under short notice.
In addition, a substantial number of manual corrections had to do with reviewers using
a different email address than the one that was used at the time of the reviewer
invitation. This is revealing of a broader issue with identifying users by email addresses
that change frequently enough to cause significant problems during the timespan of the
conference process.
The authors were then given the opportunity to rebut the reviews, to identify factual
errors, and to address the specific questions raised by the reviewers over a seven-day
rebuttal period. The exact format of the rebuttal was the object of considerable debate
among the organizers, as well as with prior organizers. At issue is to balance giving the
author the opportunity to respond completely and precisely to the reviewers, e.g., by
including graphs of experiments, while avoiding requests for completely new material
or experimental results not included in the original paper. In the end, we decided on the
two-page PDF document in conference format. Following this rebuttal period,
reviewers and ACs discussed papers at length, after which reviewers finalized their
evaluation and gave a final recommendation to the ACs. A significant percentage of the
reviewers did enter their final recommendation if it did not differ from their initial
recommendation. Given the tight schedule, we did not wait until all were entered.
After this discussion period, each paper was assigned to a second AC. The AC/paper
matching was again run through OpenReview. Again, the OpenReview team worked
quickly to implement the features specific to this process, in this case accounting for the
Preface IX
General Chairs
Horst Bischof Graz University of Technology, Austria
Daniel Cremers Technical University of Munich, Germany
Bernt Schiele Saarland University, Max Planck Institute for Informatics,
Germany
Ramin Zabih CornellNYCTech, USA
Workshop Chairs
Stefan Roth TU Darmstadt, Germany
Laura Leal-Taixé Technical University of Munich, Germany
Tutorial Chairs
Michael Bronstein Università della Svizzera Italiana, Switzerland
Laura Leal-Taixé Technical University of Munich, Germany
Website Chair
Friedrich Fraundorfer Graz University of Technology, Austria
Demo Chairs
Federico Tombari Technical University of Munich, Germany
Joerg Stueckler Technical University of Munich, Germany
XII Organization
Publicity Chair
Giovanni Maria University of Catania, Italy
Farinella
Finance Chair
Gerard Medioni Amazon, University of Southern California, USA
Publication Chairs
Albert Ali Salah Boğaziçi University, Turkey
Hamdi Dibeklioğlu Bilkent University, Turkey
Area Chairs
Kalle Åström Lund University, Sweden
Zeynep Akata University of Amsterdam, The Netherlands
Joao Barreto University of Coimbra, Portugal
Ronen Basri Weizmann Institute of Science, Israel
Dhruv Batra Georgia Tech and Facebook AI Research, USA
Serge Belongie Cornell University, USA
Rodrigo Benenson Google, Switzerland
Hakan Bilen University of Edinburgh, UK
Matthew Blaschko KU Leuven, Belgium
Edmond Boyer Inria, France
Gabriel Brostow University College London, UK
Thomas Brox University of Freiburg, Germany
Marcus Brubaker York University, Canada
Barbara Caputo Politecnico di Torino and the Italian Institute
of Technology, Italy
Tim Cootes University of Manchester, UK
Trevor Darrell University of California, Berkeley, USA
Larry Davis University of Maryland at College Park, USA
Andrew Davison Imperial College London, UK
Fernando de la Torre Carnegie Mellon University, USA
Irfan Essa GeorgiaTech, USA
Ali Farhadi University of Washington, USA
Paolo Favaro University of Bern, Switzerland
Michael Felsberg Linköping University, Sweden
Organization XIII
Poster Session
Tackling 3D ToF Artifacts Through Learning and the FLAT Dataset. . . . . . . 381
Qi Guo, Iuri Frosio, Orazio Gallo, Todd Zickler, and Jan Kautz
The structure of all these elements of the skeleton is the same. They
may be described as scaffoldings of carbonate of lime,
interpenetrated by a mesh-work of cells fused with one another, by
which the carbonate of lime has been deposited. The matrix in which
the ossicles lie is a jelly-like substance traversed by a few bands of
fibres which connect the various rods with one another. This jelly is
almost fluid in the fresh state, but when heated forms a hard
compound, possibly allied to mucin, which will turn the edge of a
razor.
The coelom is filled with a fluid, which is practically sea water with a
little albuminous matter in solution. Through the thin walls of the
papulae oxygen passes into this fluid, whence it easily reaches the
inner organs, since they are all in contact with some part of the
coelomic wall. Similarly CO2 is absorbed by the coelomic fluid from
all parts of the body, and diffuses through the papulae to the
surrounding water.
The coelom is indented by five folds, which project inwards from the
interradii. These folds are called the "interradial septa"; they are
stiffened by a calcareous deposit, which is not, however, sufficiently
dense to constitute a plate. In one of the septa the axial sinus and
stone-canal (see below) are embedded. These septa are to be
regarded as areas of lateral adhesion between the arms.
Fig. 188.—View of upper half of a specimen of Asterias rubens, which has been
split horizontally into two halves. ax.c, Axial sinus; g.d, genital duct; oe, cut
end of the oesophagus, the narrow neck of the stomach; py, pyloric sac;
py.c, pyloric caeca; r, rectum; r.c, rectal caeca; sept, interradial septum; st.c,
stomach lobe.
Soon the bivalve is forced to relax its muscles and allow the valves
to gape. The edge of the stomach is then inserted between the
valves and applied directly to the soft parts of the prey which is thus
completely digested. When the Starfish moves away nothing but the
cleaned shell is left behind. If the bivalve is small it may be
completely taken into the stomach, and the empty shell later rejected
through the mouth.
It was for a long time a puzzle in what way the bivalve was forced to
open. Schiemenz[444] has, however, shown that when the Starfish
folds itself in umbrella-like form over the prey it holds on to the
substratum by means of the tube-feet of the distal portions of the
arms, whilst, by means of the tube-feet belonging to the central
portions, it drags apart the valves by main force. He has shown
experimentally: (1) that whilst a bivalve may be able to resist a
sudden pull of 4000 grammes it will yield to a pull of 900 grammes
long continued; (2) that a Starfish can exert a pull of 1350 grammes;
(3) that a Starfish is unable to open a bivalve unless it be allowed to
raise itself into a hump, so that the pull of the central tube-feet is at
right angles to the prey. A Starfish confined between two glass plates
walked about all day carrying with it a bivalve which it was unable to
open.
A very small amount of matter passes into the rectum and escapes
by the anus, as the digestive powers of the Starfish are very
complete. The rectal caeca are lined by cells which secrete from the
coelomic fluid a brown material, in all probability an excretion, which
is got rid of by the anus.
When the meal is finished the stomach is restored to its former place
by the action of five pairs of retractor muscles, one pair of which
originates from the upper surface of the ambulacral ossicles in each
arm and extends to the wall of the stomach, where they are inserted
(Fig. 190, ret).
The tube-feet, which are at once the locomotor and the principal
sensory organs of the Starfish, are appendages of that peculiar
system of tubes known as the water-vascular system, which is
derived from a part of the coelom cut off from the rest during the
development of the animal. This system, as already mentioned,
consists of (1) a narrow "ring-canal," encircling the mouth and lying
on the inner surface of the membranous peristome; (2) a radial canal
leaving the ring-canal and running along the under surface of each
arm just above the ambulacral groove; (3) a vertical stone-canal
running from the madreporite downwards to open into the ring-canal
in the interspace between two arms. The madreporite is covered
externally by grooves lined with long cilia, and is pierced with narrow
canals of excessively fine calibre, the walls of which are also lined by
powerful cilia. Most of these narrow canals open below into a main
collecting canal, the stone-canal, but some open into a division of the
coelom termed the axial sinus, with which also the stone-canal
communicates by a lateral opening. The cavity of the stone-canal is
reduced by the outgrowth from its walls of a peculiar Y-shaped
projection, the ends being rolled on themselves in a complicated way
(Fig. 190, B). The walls of the canal consist of a layer of very long
narrow cells, which carry powerful flagella, and outside this of a crust
of calcareous deposit, which gives rigidity to the walls and has
suggested the name stone-canal.
Fig. 190.—A, view of the under half of a specimen of Asterias rubens, which has
been horizontally divided into two halves. B, enlarged view of the axial
sinus, stone-canal and genital stolon cut across. amb.oss, Ambulacral
ossicle; amp. ampullae of the tube-feet; ax.s, axial sinus; gon, gonad; g.stol,
genital stolon; marg, marginal ossicle; nerv.circ, nerve ring; oe, cut end of
oesophagus; pst, peristome; ret, retractor muscle of the stomach; sept,
interradial septum; stone c, stone-canal; T, Tiedemann's body; w.v.r, water-
vascular ring-canal.
Now at the base of the ectoderm all over the body of the Starfish
there is to be found a very fine tangle of fibrils; these are to be found
partly in connexion with small bipolar neurons lying amongst them
and partly with isolated sense-cells scattered amongst the ordinary
ectoderm cells. This nervous layer is, however, very much thickened
in certain places, so as to cause the ectoderm to project as a ridge.
One such ridge is found at the summit of each ambulacral groove
running along the whole under surface of the arm and terminating in
a cushion at the base of the median tentacle of the water-vascular
system. This ridge is called the radial nerve-cord. The five radial
nerve-cords are united by a circular cord, the nerve-ring, which
appears as a thickening on the peristome surrounding the mouth.
The sense-organs of the Starfish are chiefly the discs of the tube-
feet. Round the edges of these there is a special aggregation of
sense-cells; elsewhere, as in the skin of the back, only isolated
sense-cells are found, and it becomes impossible to speak of a
sense-organ.
(2) The "axial sinus" (ax) a wide vertical canal embedded in the
body-wall outside the stone-canal. This canal opens into the inner
perihaemal canal below; above it opens into several of the pore-
canals and into the stone-canal. The separation of the axial sinus
from the rest of the coelom is the remains of a feebly marked
metamerism in the larva.
(4) The "aboral sinus" (Fig. 192, ab), a tube embedded in the dorsal
body-wall running horizontally round the disc. The aboral sinus
surrounds the genital rachis (see p. 452) and gives off into each arm
two branches, the ends of which swell so as to surround the genital
organs. It has no connexion with the axial sinus though the contrary
has often been stated by Ludwig.[447]
Not all the cells forming the genital stolon become sexual cells. Many
degenerate and become pigment-cells, a circumstance to which the
organ owes its brown colour. In very many species of Starfish many
of the cells of the genital rachis undergo a similar degeneration, and
hence is produced the apparent aboral blood-vessel. Further, the
rachis is embedded in connective tissue which has undergone what
we may call the "lymphatic" modification, and this for want of a better
name we call the "aboral" blood-ring.
The size of the genital organs varies with the season of the year;
they are feather-shaped, and attached to the genital rachis by their
bases, but project freely into the coelom of the arm. From their great
variation in size and also from the shape of some of the cells in the
genital rachis, Hamann concludes that as each period of maturity
approaches fresh germ-cells are formed in the rachis and wander
into the genital organ and grow there in size. It is probable that the
aboral end of the genital stolon is the seat of the formation of new
germ-cells.
The genital ducts are formed by the burrowing outwards of the germ-
cells. When it is remembered that the fundamental substance of the
body-wall is semi-fluid jelly, this process will be better understood.
When the ova and spermatozoa are ripe, they are simply shed out
into the sea and fertilisation occurs there. The development is
described in Chapter XXI. The free-swimming larval period lasts
about six weeks.
In some species the fact that the number of arms exceeds 5 seems
to be connected with the power of multiplication by transverse
fission. Thus Ludwig[452] has shown that in Asterias tenuispina the
number of arms is usually 7, but sometimes 5, 6, or 8, and that in
most cases the arms are arranged in two groups—one consisting of
small arms, the other of large.
The adhesion can go so far that the animal acquires the shape of a
pentagonal disc. This is the case for instance in Culcita. The fact that
the body of this animal is really composed of adherent arms is at
once made clear when the coelom is opened. This space is found to
be divided up by inwardly projecting folds called interradial septa,
which are stiffened by calcareous deposits and represent the
conjoined adjacent walls of two arms.