Pattern Recognition ICPR International Workshops and Challenges Virtual Event January 10 15 2021 Proceedings Part IV 1st Edition Alberto Del Bimbo
Pattern Recognition ICPR International Workshops and Challenges Virtual Event January 10 15 2021 Proceedings Part IV 1st Edition Alberto Del Bimbo
https://textbookfull.com/product/natural-interaction-in-medical-
training-tools-and-applications-1st-edition-alberto-del-bimbo/
https://textbookfull.com/product/distributed-computing-and-
internet-technology-17th-international-conference-
icdcit-2021-bhubaneswar-india-january-7-10-2021-proceedings-
diganta-goswami/
https://textbookfull.com/product/pattern-recognition-and-
computer-vision-first-chinese-conference-prcv-2018-guangzhou-
china-november-23-26-2018-proceedings-part-iv-jian-huang-lai/
https://textbookfull.com/product/multimedia-modeling-20th-
anniversary-international-conference-mmm-2014-dublin-ireland-
january-6-10-2014-proceedings-part-ii-1st-edition-sema-alacam/
https://textbookfull.com/product/proceedings-of-the-11th-
international-conference-on-soft-computing-and-pattern-
recognition-socpar-2019-ajith-abraham/
https://textbookfull.com/product/advanced-information-systems-
engineering-workshops-caise-2018-international-workshops-tallinn-
estonia-june-11-15-2018-proceedings-raimundas-matulevicius/
https://textbookfull.com/product/proceedings-of-the-tenth-
international-conference-on-soft-computing-and-pattern-
recognition-socpar-2018-ana-maria-madureira/
Alberto Del Bimbo · Rita Cucchiara ·
Stan Sclaroff · Giovanni Maria Farinella ·
Tao Mei · Marco Bertini ·
Hugo Jair Escalante · Roberto Vezzani (Eds.)
LNCS 12664
Pattern Recognition
ICPR International Workshops
and Challenges
Virtual Event, January 10–15, 2021
Proceedings, Part IV
Lecture Notes in Computer Science 12664
Founding Editors
Gerhard Goos
Karlsruhe Institute of Technology, Karlsruhe, Germany
Juris Hartmanis
Cornell University, Ithaca, NY, USA
Pattern Recognition
ICPR International Workshops
and Challenges
Virtual Event, January 10–15, 2021
Proceedings, Part IV
123
Editors
Alberto Del Bimbo Rita Cucchiara
Dipartimento di Ingegneria Dipartimento di Ingegneria “Enzo Ferrari”
dell’Informazione Università di Modena e Reggio Emilia
University of Firenze Modena, Italy
Firenze, Italy
Giovanni Maria Farinella
Stan Sclaroff Dipartimento di Matematica e Informatica
Department of Computer Science University of Catania
Boston University Catania, Italy
Boston, MA, USA
Marco Bertini
Tao Mei Dipartimento di Ingegneria
Cloud & AI, JD.COM dell’Informazione
Beijing, China University of Firenze
Firenze, Italy
Hugo Jair Escalante
Computational Sciences Department Roberto Vezzani
National Institute of Astrophysics, Dipartimento di Ingegneria “Enzo Ferrari”
Optics and Electronics (INAOE) Università di Modena e Reggio Emilia
Tonantzintla, Puebla, Mexico Modena, Italy
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Foreword by General Chairs
It is with great pleasure that we welcome you to the post-proceedings of the 25th
International Conference on Pattern Recognition, ICPR2020 Virtual-Milano.
ICPR2020 stands on the shoulders of generations of pioneering pattern recognition
researchers. The first ICPR (then called IJCPR) convened in 1973 in Washington, DC,
USA, under the leadership of Dr. King-Sun Fu as the General Chair. Since that time,
the global community of pattern recognition researchers has continued to expand and
thrive, growing evermore vibrant and vital. The motto of this year’s conference was
Putting Artificial Intelligence to work on patterns. Indeed, the deep learning revolution
has its origins in the pattern recognition community – and the next generations of
revolutionary insights and ideas continue with those presented at this 25th ICPR. Thus,
it was our honor to help perpetuate this longstanding ICPR tradition to provide a lively
meeting place and open exchange for the latest pathbreaking work in pattern
recognition.
For the first time, the ICPR main conference employed a two-round review process
similar to journal submissions, with new papers allowed to be submitted in either the
first or the second round and papers submitted in the first round and not accepted
allowed to be revised and re-submitted for second round review. In the first round,
1554 new submissions were received, out of which 554 (35.6%) were accepted and 579
(37.2%) were encouraged to be revised and resubmitted. In the second round, 1696
submissions were received (496 revised and 1200 new), out of which 305 (61.4%)
of the revised submissions and 552 (46%) of the new submissions were accepted.
Overall, there were 3250 submissions in total, and 1411 were accepted, out of which
144 (4.4%) were included in the main conference program as orals and 1263 (38.8%)
as posters (4 papers were withdrawn after acceptance). We had the largest ICPR
conference ever, with the most submitted papers and the most selective acceptance
rates ever for ICPR, attesting both the increased interest in presenting research results at
ICPR and the high scientific quality of work accepted for presentation at the
conference.
We were honored to feature seven exceptional Keynotes in the program of the
ICPR2020 main conference: David Doermann (Professor at the University at Buffalo),
Pietro Perona (Professor at the California Institute of Technology and Amazon Fellow
vi Foreword by General Chairs
at Amazon Web Services), Mihaela van der Schaar (Professor at the University of
Cambridge and a Turing Fellow at The Alan Turing Institute in London), Max Welling
(Professor at the University of Amsterdam and VP of Technologies at Qualcomm),
Ching Yee Suen (Professor at Concordia University) who was presented with the IAPR
2020 King-Sun Fu Prize, Maja Pantic (Professor at Imperial College UK and AI
Scientific Research Lead at Facebook Research) who was presented with the IAPR
2020 Maria Petrou Prize, and Abhinav Gupta (Professor at Carnegie Mellon University
and Research Manager at Facebook AI Research) who was presented with the IAPR
2020 J.K. Aggarwal Prize. Several best paper prizes were also announced and awarded,
including the Piero Zamperoni Award for the best paper authored by a student, the
BIRPA Best Industry Related Paper Award, and Best Paper Awards for each of the five
tracks of the ICPR2020 main conference.
The five tracks of the ICPR2020 main conference were: (1) Artificial Intelligence,
Machine Learning for Pattern Analysis, (2) Biometrics, Human Analysis and Behavior
Understanding, (3) Computer Vision, Robotics and Intelligent Systems, (4) Document
and Media Analysis, and (5) Image and Signal Processing. The best papers presented at
the main conference had the opportunity for publication in expanded format in journal
special issues of IET Biometrics (tracks 2 and 3), Computer Vision and Image
Understanding (tracks 1 and 2), Machine Vision and Applications (tracks 2 and 3),
Multimedia Tools and Applications (tracks 4 and 5), Pattern Recognition Letters
(tracks 1, 2, 3 and 4), or IEEE Trans. on Biometrics, Behavior, and Identity Science
(tracks 2 and 3).
In addition to the main conference, the ICPR2020 program offered workshops and
tutorials, along with a broad range of cutting-edge industrial demos, challenge sessions,
and panels. The virtual ICPR2020 conference was interactive, with real-time
live-streamed sessions, including live talks, poster presentations, exhibitions, demos,
Q&A, panels, meetups, and discussions – all hosted on the Underline virtual confer-
ence platform.
The ICPR2020 conference was originally scheduled to convene in Milano, which is
one of the most beautiful cities of Italy for art, culture, lifestyle – and more. The city
has so much to offer! With the need to go virtual, ICPR2020 included interactive
virtual tours of Milano during the conference coffee breaks, which we hoped would
introduce attendees to this wonderful city, and perhaps even entice them to visit Milano
once international travel becomes possible again.
The success of such a large conference would not have been possible without the
help of many people. We deeply appreciate the vision, commitment, and leadership
of the ICPR2020 Program Chairs: Kim Boyer, Brian C. Lovell, Marcello Pelillo, Nicu
Sebe, René Vidal, and Jingyi Yu. Our heartfelt gratitude also goes to the rest of the
main conference organizing team, including the Track and Area Chairs, who all gen-
erously devoted their precious time in conducting the review process and in preparing
the program, and the reviewers, who carefully evaluated the submitted papers and
provided invaluable feedback to the authors. This time their effort was considerably
higher given that many of them reviewed for both reviewing rounds. We also want to
acknowledge the efforts of the conference committee, including the Challenge Chairs,
Demo and Exhibit Chairs, Local Chairs, Financial Chairs, Publication Chair, Tutorial
Chairs, Web Chairs, Women in ICPR Chairs, and Workshop Chairs. Many thanks,
also, for the efforts of the dedicated staff who performed the crucially important work
Foreword by General Chairs vii
behind the scenes, including the members of the ICPR2020 Organizing Secretariat.
Finally, we are grateful to the conference sponsors for their generous support of the
ICPR2020 conference.
We hope everyone had an enjoyable and productive ICPR2020 conference.
Rita Cucchiara
Alberto Del Bimbo
Stan Sclaroff
Preface
Competitions are effective means for rapidly solving problems and advancing the state
of the art. Organizers identify a problem of practical or scientific relevance and release
it to the community. In this way the whole community can contribute to the solution of
high-impact problems while having fun. This part of the proceedings compiles the best
of the competitions track of the 25th International Conference on Pattern Recognition
(ICPR).
Eight challenges were part of the track, covering a wide variety of fields and
applications, all of this within the scope of ICPR. In every challenge organizers
released data, and provided a platform for evaluation. The top-ranked participants were
invited to submit papers for this volume. Likewise, organizers themselves wrote articles
summarizing the design, organization and results of competitions. Submissions were
subject to a standard review process carried out by the organizers of each competition.
Papers associated with seven out the eight competitions are included in this volume,
thus making it a representative compilation of what happened in the ICPR challenges.
We are immensely grateful to the organizers and participants of the ICPR 2020
challenges for their efforts and dedication to make the competition track a success. We
hope the readers of this volume enjoy it as much as we have.
General Chairs
Rita Cucchiara Univ. of Modena and Reggio Emilia, Italy
Alberto Del Bimbo Univ. of Florence, Italy
Stan Sclaroff Boston Univ., USA
Program Chairs
Kim Boyer Univ. at Albany, USA
Brian C. Lovell Univ. of Queensland, Australia
Marcello Pelillo Univ. Ca’ Foscari Venezia, Italy
Nicu Sebe Univ. of Trento, Italy
René Vidal Johns Hopkins Univ., USA
Jingyi Yu ShanghaiTech Univ., China
Workshop Chairs
Giovanni Maria Farinella Univ. of Catania, Italy
Tao Mei JD.COM, China
Challenge Chairs
Marco Bertini Univ. of Florence, Italy
Hugo Jair Escalante INAOE and CINVESTAV National Polytechnic
Institute of Mexico, Mexico
Publication Chair
Roberto Vezzani Univ. of Modena and Reggio Emilia, Italy
Tutorial Chairs
Vittorio Murino Univ. of Verona, Italy
Sudeep Sarkar Univ. of South Florida, USA
Web Chair
Andrea Ferracani Univ. of Florence, Italy
Local Chairs
Matteo Matteucci Politecnico di Milano, Italy
Paolo Napoletano Univ. of Milano-Bicocca, Italy
Financial Chairs
Cristiana Fiandra The Office srl, Italy
Vittorio Murino Univ. of Verona, Italy
Contents – Part IV
Bias from the Wild Industry 4.0: Are We Really Classifying the Quality
or Shotgun Series? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 637
Riccardo Rosati, Luca Romeo, Gianalberto Cecchini, Flavio Tonetto,
Luca Perugini, Luca Ruggeri, Paolo Viti, and Emanuele Frontoni
Workshop Description
The ubiquitous surveillance cameras are generating huge amount of videos. Automatic
video content analysis and recognition are thus desirable for effective utilization of
those data. Fine-Grained Visual Recognition and Re-Identification (FGVRID) aims to
accurately identify visual objects and match re-appearing targets, e.g., persons and
vehicles from a large set of images and videos. It has the potential to offer an
unprecedented possibility for intelligent video processing and analysis, as well as to
explore the promising applications on public security.
The FGVRID workshop wishes to bring together researchers from fine-grained
visual categorization, as well as person/ vehicle ReID communities, and to foster
discussions and exchange of ideas between them. FGVRID is not a traditional search or
classification task due to its goal of accurately identifying visual objects. First, proper
detection algorithms should be designed to locate objects and their parts in videos
before proceeding to the identification step. Second, the visual appearance of an object
is easily affected by many factors like viewpoint changes and camera parameter dif-
ferences, etc. Third, annotating the fine-grained identity or category cues is expensive
and time consuming. Finally, to cope with the large-scale data, scalable indexing or
feature coding algorithms should be designed to ensure the online recognition effi-
ciency. Aiming to seek novel solutions and possibilities in FGVRID, this workshop
will have in-depth discussions on those issues and aims to go beyond toy datasets and
small-scale algorithms. A total of 12 submissions were received and after a single-blind
reviewing process including 2-3 reviewers per paper, 7 papers were accepted. The
acceptance rate was 58%.
The topics of FGVRID were equitably represented with 7 presentations focusing on
semi-supervised learning, effective video representations, fine-grained classification,
new datasets for fine-grained visual recognition. Four speakers were invited to give
talks in the workshop. Each speaker gave 40mins talk on fine-grained visual recog-
nition and person/vehicle ReID or related topics.
Organization
General Chairs
Program Committee
1 Introduction
The State of the Art (SoA) deep learning methods in computer vision achieve
high accuracy by leveraging large, diverse and correctly annotated datasets, with
the most detailed annotations desired being at a pixel level. For problems like
crowd event detection, pose estimation, person detection, recognition, segmenta-
tion, re-identification, tracking, activity recognition, the production of detailed
Funded under the H2020 project MindSpaces, Grant number # 825079.
c Springer Nature Switzerland AG 2021
A. Del Bimbo et al. (Eds.): ICPR 2020 Workshops, LNCS 12664, pp. 5–19, 2021.
https://doi.org/10.1007/978-3-030-68799-1_1
6 R. Montulet and A. Briassouli
2 Related Work
The role of synthetic datasets as supplements to existing training data, or as
data in and of themselves, is receiving increased attention [36,37,43,44], as deep
learning requires extensive high quality annotated data to perform well, which
is not always easy to obtain in the real world. To this end, various synthetic
datasets have been recently generated to solve different computer vision prob-
lems, with the works in [6,25,37,38,40,49] all using synthetic data. Several of
them, namely [25,37,38,49] use the GTA V Rockstar Advanced Game Engine
(RAGE), similarly to our work, but focusing on different computer vision prob-
lems. The reason for choosing RAGE is the high quality of the resulting graphics,
as well as the policy of the game engine, which allows for non-commercial use
of its footage [2]. Data generation tools and datasets that use virtual worlds to
generate annotated image datasets are described below.
A wide range of synthetically generated data has been produced for com-
puter vision problems https://github.com/unrealcv/synthetic-computer-vision,
but not for the application that we are examining, namely crowd event detec-
tion. Moreover, unlike our dataset, existing ones do not offer the flexibility to be
used for several other applications, from recognition of individual or group activ-
ities/interactions in a variety of scenarios and environments, to person tracking,
segmentation re-identification and others, as detailed below.
Synthetic Datasets for Optical Flow. Synthetic datasets have been used
to develop accurate optical flow algorithms since 1987 [21], with Yosemite [7]
(1994) being one of the most widely used. In [30], what makes a good synthetic
dataset is described, with an extensive overview of existing synthetic optical
flow benchmarking datasets, including recent ones like Flying Chairs [15], and
SYNTHIA [39]. It should be noted that the SoA optical flow Flownet2 [24] used
Densely Annotated Photorealistic Virtual Dataset 9
in this work has also been developed using synthetically generated data, which
allowed it to outperform the SoA.
Synthetic Human, Crowd Datasets. Previous efforts on synthetic data gen-
eration for crowd simulation [5,31] focus on crowd group dynamics, but not on
the quality of the graphics, making them less appropriate for deep learning, which
requires large amounts of high quality annotated data. Recently, a dataset and
tool for crowd counting was published [48], which only generates crowd images,
and not crowd videos, nor crowd event scenarios.
In [12], a dataset for human activity recognition has been generated, but with
Unity, contains one person per activity, and no abnormal crowd events. Motion
tracking and activity recognition can take place using the synthetic data in [19],
however it features only one person per frame, and has a blank background.
Human segmentation and depth estimation datasets have also been synthet-
ically generated recently [47], based on motion capture data. However, they use
Blender [11], and comprise of single person images rather than continuous video.
A large scale synthetic image dataset of images of street scenes with dense seman-
tic segmentation maps, generated by the Unity game engine, is SYNTHIA. It has
been used for semantic image segmentation, image-to-image translation [23,28]
and adversarial domain adaptation [22], among others.
Our tool generates densely annotated crowds and events, but can also be
used for the generation of individual or small group activities, tracking, pose,
segmentation, providing solutions for a wider range of computer vision problems
than existing datasets. At the same time, it provides densely annotated bench-
marking data for abnormal crowd event detection, for which existing real-world
datasets have been limited in size, quality and amount of events (see Sect. 2.3).
annotations. However, this dataset does not contain crowd events or abnormal
individual/crowd behaviors, as it focuses on abnormalities related to street scenes
and rules, such as pedestrians crossing illegally or bicycles on sidewalks.
Our dataset features significantly more frames than most of the above
datasets, with the exception of StreetScene, however the number of frames in
our dataset can be directly increased by using our tool. Our dataset also features
a wide range of weather conditions and environments, as well as events related
to abnormalities in crowd behavior and motion, rather than anomalies related
to appearance changes in one frame (as in UCSD, UMN). Thus it is more appro-
priate for detecting abnormalities in videos, in the behavior of crowds but also
individuals. It does not contain annotations only for events, but also for person
segmentation, tracking, re-identification, pose classification and more, detailed
in Sect. 3.2. Its graphics are of high quality, as they have been generated with
the GTA V engine [14], which can produce a wide range of high quality, photo-
realistic scenes.
Fig. 1. The distribution of group behaviours, weathers, frames per second used to
render the videos and of the times during which the videos were recorded.
Fig. 2. Our frame annotations: from left to right, the bounding boxes of people along
with their ID and distance to the camera, the bone coordinates of the people in the
scene, the trajectories people have walked, the pixel-wise segmentation of people, and
lastly pixel-wise depth information relative to the camera plane.
Now that the scene is set up, the timescale is set to 1.0, and the game starts
to render the scene. Meanwhile, the script saves images and annotations, which
include RGB, depth, segmentation information, person pose information (bone
12 R. Montulet and A. Briassouli
locations), bounding boxes, and person IDs. Thus, the post-processing script
can find trajectories, tighten bounding boxes, segmentations, and depth esti-
mates. Figure 2 shows a sample synthetically generated frame using our method,
with annotations corresponding to person detection (bounding boxes), bones,
trajectories, person segmentation and depth maps.
Hän oli lakitta päin. Lakki oli viereisellä tuolilla rutistuneena. Hänen
päänsä oli kaunismuotoinen ja hänen lyhyeksi leikattu tumma
tukkansa oli miehekäs ja kiiltävä. Hänen kauniilla valkoisella
otsallaan riippui muutamia hien suoristansa hiuksia. Josephine
ajatteli, että se oli yksi syy, jonka vuoksi hän oli päättänyt, että
Lattimer oli häikäilemättömän välinpitämätön. Sivistynyt ja
huolellinen mies olisi pyyhkäissyt hiukset hikineen päivineen
otsaltaan naisen seurassa.
Loput hänen häikäilemättömyydestään kuvastui hänen silmissään,
suun viivoissa ja hänen voimakkaassa leuassaan.
Mutta nyt hän alkoi tulla levottomaksi sen huomion johdosta, että
hänen täytyi tehdä erinäisiä selvityksiä matkastaan palattuaan
Triangle L:ään. Oli selvää, että häntä kaivattaisiin, samoin kuin
Chesterfieldiä ja sitä hevosta, jolla Artwell oli ratsastanut. Hän mietti
harmillista asemaa, johon hän oli itsensä saattanut. Joko hänen
täytyy tehdä tiliä yöllisistä toimistaan Bettylle taikka sitten valehdella,
jonka hän luonnollisesti aikoikin tehdä, estääkseen Brannonia
ratsastamasta suoraapäätä Artwellin niskaan. Ja hän punastui ja
ilmaisi levottomuutensa tarttumalla hermostuneesti tuolin molempiin
nojiin. "Kas niin", sanoi Lattimer heti, "tehän olette aivan
suunniltanne. Mutta teillä onkin ollut kova työ tänä yönä. Istukaa
paikallanne muutama minuutti, sillä aikaa kun koetan saada kokoon
jotakin syötävää."
"Jos olette varma Mrs Whitmanista, niin —", alkoi Josephine. Hän
oli oikein innokas nyt ja ihaili Lattimerin kekseliäisyyttä.
"Joutavia!" keskeytti Lattimer hänet, hymyillen varmuudessaan.
"Tuntuuko paremmalle?"
"Paljon paremmalle."
Vaihtamistaan silmäyksistä saattoi kumpainenkin lukea toistensa
ajatukset. Josephine näki selvästi, että Lattimeria oli huvittanut
hänen hämminkinsä. Hän luki selvästi hänen ilmeestään: "Kas niin,
sinulla ei ollut mitään syytä peljätä minua loppujen lopuksi, vai mitä?"
"Te tekasitte pienen jutun, Miss Hamilton", sanoi hän. "En ole
mikään ajatusten lukija, mutta te annoitte äsken itsenne ilmi.
Nähkääs, minä tunsin Callahanin ja Callahan ei koskaan olisi tehnyt
sitä, mistä te häntä syytitte. Luonnollisesti teidän täytyi esittää jotakin
raskauttavaa kun kerran tahdoitte pelastaa Brannonin. En vain
ymmärrä miksi te tahdoitte pelastaa Brannonin Callahanin maineen
kustannuksella. Jos kerran Brannon on tehnyt murhan, niin on
hänessä kyllin miestä kantamaan sen seuraukset ilman mitään
sellaisia juttuja, joita te esititte uskottaviksi."
"Te ette, näen mä, ole kuullut siitä. No niin, toistaiseksi ei ole vielä
mitään todistettua. Mutta ympäristön karjamiehillä on varsin hyvin
perusteltu epäluulo, että näillä seuduilla toimii hyvin järjestetty
hevosvarasjoukkue. Sangen monta miestä epäillään, ehkäpä
minuakin. Les Artwell on toinen, vaikka luulen, että hän on yhtä
syytön kuin minäkin."
Hän oli sitonut Artwellin haavan, joka paistoi avonaisen paidan alta
rinnassa. Artwell oli taas tajuissaan, vaikkakin hänen silmänsä olivat
kuumeesta kiiltävät ja kasvonsa punakat.
"Arvaan, että kun mies tekee tuollaisia tekoja, niin on tyttö silloin
saanut hänen päänsä täysin pyörälle", lisäsi hän lopuksi.
"Sekin oli liikaa" sanoi Lattimer. Hänen äänensä oli kuiva ja selvä,
mutta hillitty ja siinä oli niin selvä uhka, että Artwellinkin kasvot
kalpenivat. Denverin saattoi Lattimerin uhkaava käytös aivan
hervottomaksi ja hän seisoi liikkumattomana, pelokkaana ja huulet
lerpallaan.
Kahdeksastoista luku.
Hän oli satuloinut sen, heittänyt toisen ohjaksen sen pään yli ja oli
juuri sulkemassa aitauksen porttia, kun hän katsahti talliin päin.
Kiirehtimättä sulki hän portin, hyväili kädellään mustaa hevostaan ja
meni tallia kohti.
Hän meni noutamaan Bettyn hevosen ja satuloi sen sillä aikaa kun
Betty vaihtoi pukua päärakennuksessa.
"No niin, näkemiin sitten", sanoi hän. Hän ratsasti pari kymmentä
askelta, pysäytti hevosensa ja huusi olkansa yli puoleksi leikillään:
"Sehän on vaan tyttö", hymyili Brannon. "Ei nyt ole kysymys Starin
miehistä."