SHAPEATLAS_20_KUL - Computer Graphics and Knowledge

advertisement
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
Small or medium-scale focused research project (STREP)
ICT Call 9
FP7-ICT-2011-9
SHAPEATLAS
Type of project:
Small or medium scale focused research project (STREP)
Date of preparation:
April 17th, 2012
Work programme objective addressed:
Objective ICT-2011.8.2 ICT for access to cultural resources
Name of the coordinating person:
Torsten Ullrich
e-mail:
torsten.ullrich@fraunhofer.at
fax:
+43 316 873 105404
Participant no.
1 (Coordinator)
2
3
4
5
Participant organisation name
Fraunhofer Austria Research GmbH
University of Brighton
Technische Universität Graz
Katholieke Universiteit Leuven
Brunswick Town Charitable Trust
Part. short name
FhA
UOB
TUG
KUL
BTCT
Country
Austria
United Kingdom
Austria
Belgium
United Kingdom
Proposal abstract
The SHAPEATLAS project will advance the state of the art in the access to cultural resources, in particular
to large collections of similar 3D shapes, for public, research, and education. Searching in collections is a
delicate task. Museums typically collect many objects of the same kind, e.g., coins, swords, vases, statues,
etc. Art historians are typically interested in just a few subtle properties of a given shape which they use for
various hypotheses, e.g., about the era, the location of origin, about cultural influences, or the diffusion of
technology. There are many examples of shape typologies, e.g., of columns, of helmets, or of amphorae
(Dressel typology with 66 types). Similarity measures that are based on machine learning treat shape
description and classification as a purely statistical problem (clustering of feature vectors etc.). Such methods
fail in distinguishing subtle differences of objects in a collection. And, more importantly, there is no way for
users to explicitly specify what to look for. Small shape differences can lead to a very different classification.
The SHAPEATLAS project will introduce an innovative concept to solve the problem: parametric shape
maps. A shape map is basically a user-defined shape measurement tool. It can adapt to a given object
(scanned sword) to measure its characteristic parameters (blade length and width). Since a shape can be
measured in different ways, several shape maps can exist in parallel. SHAPEATLAS will investigate
methods to define and to apply shape maps interactively, but also automatically, e.g., to classify a larger
collection overnight. Novel searching and browsing tools will allow searching for statistical and for
parametric shape properties, thus combining both approaches. The project results will be widely applicable,
which will be demonstrated with the use case of building restoration (large number of moulds) and museum
collection management.
Proposal Part B: page 1
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
Table of Contents
1
2
Scientific and/or technical quality, relevant to the topics addressed by the call .......................... 3
1.1
Concept and objectives .......................................................................................................... 3
1.2
Progress beyond the state-of-the-art .................................................................................... 10
1.3
S/T methodology and associated work plan ........................................................................ 12
1.3.1
Work package list......................................................................................................... 12
1.3.2
List of deliverables ....................................................................................................... 12
1.3.3
List of milestones ......................................................................................................... 13
1.3.4
Work package description ............................................................................................ 14
1.3.5
Summary of effort ........................................................................................................ 22
Implementation .......................................................................................................................... 23
2.1
2.1.1
Management roles ........................................................................................................ 23
2.1.2
Decision-making bodies and mechanisms ................................................................... 23
2.1.3
Reporting mechanisms ................................................................................................. 23
2.1.4
Meetings, conferencing and communication tools ...................................................... 24
2.1.5
Knowledge management and intellectual property rights ............................................ 24
2.1.6
Quality assurance and control ...................................................................................... 24
2.1.7
Conflict management ................................................................................................... 24
2.2
3
Management structure and procedures ................................................................................ 23
Individual participants ........................................................................................................ 25
2.2.1
Fraunhofer Austria Research GmbH............................................................................ 25
2.2.2
University of Brighton ................................................................................................. 26
2.2.3
Technische Universität Graz ........................................................................................ 27
2.2.4
Katholieke Universiteit Leuven ................................................................................... 28
2.2.5
Brunswick Town Charitable Trust ............................................................................... 29
2.3
Consortium as a whole ....................................................................................................... 29
2.4
Resources to be committed ................................................................................................. 29
Impact......................................................................................................................................... 30
3.1
Expected impacts listed in the work programme ................................................................ 30
3.2 Dissemination and/or exploitation of project results, and management of intellectual
property .......................................................................................................................................... 30
4
Ethical Issues.............................................................................................................................. 31
5
References .................................................................................................................................. 32
Proposal Part B: page 2
FP7-ICT-2011-9
18/01/12 v1
1
1.1
STREP proposal
SHAPEATLAS
Scientific and/or technical quality, relevant to the topics addressed by the call
Concept and objectives
The SHAPEATLAS project will greatly advance the state of the art in the access to cultural
resources, especially to very large collections of 3D shapes, for public, research, and education. The
availability of 3D-representations of cultural artefacts must become a standard in all sectors that are
concerned with cultural artefacts. In the near future, whenever a specific cultural object is
mentioned, there will be the expectation that a 3D-representation of it is available; that various other
similar objects can be obtained through 3D searching and browsing; and that subtle shape variations
are detected and highlighted. Especially important is the case of collections of many similar objects.
It is rarely the case that a museum will host only one object of a kind. Instead, museums typically
collect many coins, or many swords, or are specialized on vases, or statues, or watches, or fire arms,
etc. Automatic shape classification and mark-up are not applicable in this case because items in a
collection can often be described best in terms of their characteristic shape parameters. This
parameter set, however, is dependent of the shape class under consideration, so there has to be a
way for the human operator to define a “parametric shape template” that explains to the computer
how to measure the characteristic parameters. This template then allows for automatic classification
of the shapes in the collection, thereby obtaining a more targeted shape database – or merely a
digital shape library – with much richer semantics. The semantic shape classification is the basis for
many novel applications, including collection management, outlier detection, and even statistical
data mining.
Problem Statement
The availability of collections of real-world artefacts as three-dimensional digital assets will enable
a wide range of novel methods from cultural and art history, over collections management, to
statistical comparative studies. Easier access to collections will stimulate the exchange of research
results and facilitate a broader, more complete, and more precise historical knowledge. All this is
possible, however, only with rich high-level shape semantics. “Semantics is the key” describes the
difference between a simple database and a digital library: Shape is not just treated as data (e.g.,
scanned triangle meshes), but the digital library must also contain a formalized notion of the
“meaning” of the shape data that goes beyond, e.g., statistical shape descriptors. – In order to meet
this goal, we have identified four main problems.
The first problem is the cost-efficient 3D-acquisition of a collection of real-world shapes. 3D
scanning has traditionally been an expert job. 3D scan data preferably requires manipulation and
cleanup of skilled people. These days 3D-acquisition is greatly facilitated today by
photogrammetric methods. Especially in the Cultural Heritage (CH) sector, these methods have
enormous potential and may very well revolutionize archaeological documentation: The intention is
to go for 3D acquisition methods that are user-friendly and portable to be used in a very friendly
way. Having for instance image based 3D acquisition is much easier and cheaper than operating a
3D scanner and would allow the acquisition done by non-experts.
In particular; methodologies purely based on photo sequences can cover an object faster in a more
complete way, not having to rely on bulky acquisition setups; CH objects will typically exhibit a
rich texture and are typically well suited for photogrammetric methods; image registration and
sparse point cloud generation are completely automatic; and even dense matching and 3D
reconstruction of individual image sequences can be largely automated. For larger and more
complete objects, however, a large number of image sequences must be processed. One problem is
scalability, since very large amounts of data must be handled routinely; another is interactivity and
quality control, to overcome the problem of detecting too late that parts are missing; and the third is
sustainability in the sense that new additional image data can still be integrated later on. In this
project it is believed that having “shape” knowledge on certain types of CR specimen, can also
benefit the guidance of the acquisition on the one hand, and the final creation of the model. It
Proposal Part B: page 3
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
should be noted that the approach will not be restricted to purely photographic methods, also
methodologies involving structured light and or photometric based methods can and will be tuned
towards user-friendly acquisition.
The second problem is to describe the typology of a collection of similar objects, like amphorae or
swords. Automatic solutions exist for shape search and shape matching; this is a well-established
research field. These approaches, however, are not applicable in the case of collections because the
differences in shape are too subtle. Automatic shape classification is very effective for
discriminating shape classes (amphora vs. sword). With shape collections, however, small intraclass shape differences can make a large difference in classification. Subtle shape differences are
just the information that art historians are typically interested in most because they allow drawing
various conclusions, e.g., about the era, the location of origin, about cultural influences, or the
diffusion of technology. There are many examples of classical typologies including column types
(Doric, Ionic, Corinthian), helmet types (Attic, Thracian, Corinthian, Illyrian, Phrygian), and also
amphorae (about 66 types, systematic classification by Heinrich Dressel and others). So the task is
to find user friendly ways for defining “parametric shape templates”: Given a digital artifact, an art
historian must be able to describe which parts of the shape are more important for a particular
classification than others – keeping in mind that any classification is just an interpretation, i.e.,
several classification schemes may exist in parallel.
The third problem concerns the documentation of large collections of shapes. Many museums
suffer from a considerable backlog in documentation. Museum archives are full of objects that still
need to be described, but this tedious and costly. Every collection item must be described,
classified, photographed, and measured. Putting objects on display additionally requires cleaning
and restoration; but before deciding what is displayed, it is essential to document the collection.
Again, photogrammetric methods have a great potential: Once a 3D reconstruction from one or
more photo sequences is available, the 3D shape can be computed. Measurements could then be
taken from the digital shape, using appropriate easy to use 3D-measurement tools. Once a
measurement procedure for one object is defined, also similar objects can be measured using the
same procedure. An obvious extension would be to carry out these measurements automatically for
all objects that belong to the same shape class, thus progressing from semi-automatic to automatic
shape measurement.
And finally, the fourth problem is searching and browsing a large collection of digital artefacts.
Conventional shape search uses an input shape and produces all similar shapes in the database as
output. This is not very reasonable for a collection, where all shapes are similar. Conventional
similarity measures are based on shape descriptors and are computed by comparing feature vectors.
This is indispensable for managing larger collections, as this allows fast navigation through
different shapes. An important extension, however, would be to complement it by the parametric
search based on shape measurements. So the user could use similarity-based search to interactively
refine the search results, and at some point switch to parameter/measurement-based search.
Parametric modelling and the inverse problem, shape fitting
3D modelling of a collection of complex shapes can be greatly improved in efficiency by
parametric modelling. Conventional 3D modelling systems (SketchUp, AutoCAD, 3DStudioMax,
Maya) use a forward-modelling style where a shape is interactively edited and changed until it
matches the specification; result of the modelling process is a set of triangles or NURBS patches.
Parametric high-end modelling systems (Pro/Engineer, CATIA) internally use a parametric
description that allows for parameter changes to easily create variants of a model, e.g., when the
specifications change.
Generative modelling goes one step further in that a shape is described as a sequence of shape
modelling operations. This reflects a paradigm change from objects to operations, and permits for
Proposal Part B: page 4
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
creating libraries of parametric shapes that can be nested. Fig. X shows an example of a
hierarchically structured Gothic windows where (i) the defined regions (circle, fillets and subwindows) can be instantiated with different shapes from a library (top left), and (ii) the hierarchy
can be further refined by inserting sub-shapes that again define new regions (bottom left). This
example uses the GML shape description language for generative models developed by partner
TUG (Havemann, 2005).
Fig. X: Generative models of Gothic window tracery from [Hav05]. The generative description is
not only parametric, but also hierarchical; different sorts of sub-shapes can be inserted from a
library.
While this solves the forward modelling problem for collections, SHAPEATLAS is confronted with
the inverse problem, i.e., shape analysis, and not shape modelling. One way to analyse a shape,
however, is to use an analysis-by-synthesis approach. This means to re-generate the scanned target
shape by finding the appropriate shape parameters for it, as is illustrated in Fig. X. Although this
seems to be an elegant way to solve the analysis problem, it turns out that finding the parameters for
non-trivial shapes numerically is extremely hard, as is explained in the next section.
Abandoning fully automatic fitting of complex parametric shapes
Fitting to a mesh a detailed parametric template model with many degrees of parametric variability
implies a high development effort and computational costs. The conventional approach is to search
for optimal parameter configurations in a piecewise linear, high-dimensional, curved shape
parameter space. This naturally leads to non-convex constrained optimization problems that are
difficult to solve and can easily run into undesired local minima. Random sampling and multiresolution approaches can help, but processing times remain prohibitive. Also unsolved is the sheer
amount of memory required.
Proposal Part B: page 5
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
Fig. X: Inverse generative modelling. Given a scanned shape (left) and a parametric shape (right),
the goal is to find the appropriate set of parameters that gives the optimal match to the scanned
shape.
Another important problem is that CH artefacts are often delicate and elaborate pieces of
craftsmanship that simply defy a notion of the industrial construction process for which generative
modelling approaches work so well. Virtually all objects produced in the pre-industrial era have
shape irregularities and imperfections that are difficult to capture parametrically. Automatic fitting
is almost impossible if corresponding parameters are not incorporated in the template model
beforehand. Potentially, though, methods from statistical shape analysis could help to determine the
variability of a shape class in order to assist the operator in producing a suitable parametric model.
Towards shape semantics from shape measurements
SHAPEATLAS will therefore pursue a different approach that is conceptually novel, easily
extensible and, most importantly, directly useful for CH experts. The idea is to develop a fast and
lightweight interactive shape measurement tool. It proceeds by user-assisted fitting to a given digital
CH artefact a structured shape detector that is composed of a hierarchy of shape detector primitives.
The result is semantic information, namely a rich set of object-adapted measurements. This
addresses a practical problem in CH, the high cost of measuring and characterizing artefacts: The
2009 inventory catalogue of the Museum of Archaeology in Castle Eggenberg, Graz, lists for a
family of Slovenian military helmets from 4th to 1st century BC the following parameters: height,
diameter of inner and outer rim, and material thickness. – The complexity of measuring such values
is one reason for the large backlog of unprocessed artefacts in museum archives. Humans are good
at determining and distinguishing shape categories, while computers are well-equipped for precise
and fast numerical fitting. Fitting works extremely fast if it starts close to an optimal parameter set
with respect to alignment and shape parameters.
A shape map is a user-defined model for 3D shape measurement and classification. Technically, it
is a hierarchically structured complex shape detector that consists of simpler shape detectors, down
to primitive detectors. The goal is to provide a library of primitive shape detectors and a tool to
arrange them hierarchically. The main advantage of this approach is that shape features are
disambiguated: Small features can make large semantic differences, e.g., decorative tin plates may
easily be mistaken for shields. Subtle shape features are overlooked by unsupervised fitting
approaches, or ignored because no parameters exist for them in the model. Hence, the user supports
the system by making explicit which shape feature is important and which is not. This leads to a
two-stage process:
1. Definition of a shape map: The user starts by loading a triangle mesh of a prototype object,
e.g., a helmet. As first primitive detector he selects an empty-sphere detector for the head
Proposal Part B: page 6
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
space and roughly aligns it. The fitting procedure binds its four float parameters (origin and
radius). Next, a below-plane is selected, roughly aligned, and fitted to the base of the
helmet; then two cylinder detectors yield the radii of inner and outer rim. To fix the pose,
the rotation around the central axis is computed by applying either a symmetry-plane or a
SVD decomposition of the mesh vertices. Finally, the height is measured by inserting
another fitting plane from above, which is constrained to be parallel to the below-plane.
Once such a hierarchical detector is defined, it is stored in a shape map library. As byproduct this leads to a shape taxonomy as shape maps can later be successively refined and
specialized to match sub-classes of shapes in the collection.
2. Instantiating a shape map: The user starts by loading the triangle mesh of a newly acquired
helmet. He selects from the shape map library the helmet map and roughly aligns it. The
fitting process starts immediately, successively “snapping” more and more shape detectors
to the mesh, and it stops whenever further user assistance is required. Since this is the costly
part, the user-assisted fitting process must be very fast. We target a time budget of three to
five minutes max. This will require finding appropriate “snapping strategies”.
List of primitive shape detectors
Shape maps are combined from a well-set of primitive shape detectors. Thus, they constitute the
basis of the method. These primitives must be designed in such a way that they are both meaningful
and easily applicable. Therefore they are not “numerical magic”, but they have an intuitive meaning
and behave in a robust and predictable way. The shape detectors determine the “semantic
vocabulary” for describing and analysing shape, and must therefore be expressive enough to allow
“formulating” what constitutes the shapes of a collection. The set should therefore contain at least:








Min-max probes: planar rectangle or disc with a linear spring. It is positioned in space and
moves in normal direction in a definable range until it is blocked by the target mesh. So the
detector might "press" from the outside inwards or from inside outwards.
Segmentation volume (box/sphere/cylinder): To force fitting algorithms to use only selected
portions of the target mesh. The shape is segmented into regions to apply there other
detectors.
Fitting volume (box/sphere/cylinder): Classical fitting of a shape primitive using the onesided Hausdorff distance as fitness measure in a RANSAC approach inside a segmentation
volume.
Crease, Edge and Corner detectors: These are shape features that often carry important
semantic information. They can also be configured to search for rectangular configurations
(right angles), and for convex or concave cases.
Slippage Analysis (profile swept along line or circle segment): When applied to a surface
point, this detector looks for a high-curvature profile in a normal plane that can be swept
along a low-curvature line or circular arc. The result is both the profile curve and the spine
of the sweep. As the spine length is maximized this can, e.g., yield the height of a column.
Orientation plane: planar rectangle or disc that supports the notion of principal planarity. A
coarse brick wall for example might be roughly planar but noisy, have fissures, holes, etc. A
weighted least squares fit incorporating local first-order information can compute a
preferred normal direction. Alternative methods are ICP or RANSAC plane fitting.
Symmetry plane: A plane is roughly aligned inside a fitting volume, the four plane
parameters are optimized using a RANSAC approach followed by mirroring and
approximate Hausdorff distance.
Principal axes: The least squares normal equations can be solved by singular value
decomposition (SVD). This yields a coordinate system for initial alignment of subsequent
detectors.
Proposal Part B: page 7
FP7-ICT-2011-9
18/01/12 v1

STREP proposal
SHAPEATLAS
Similarity search region: This establishes the link to the conventional shape-based search.
The user can specify a region of the surface to be indexed using a statistical shape similarity
measure. This way also shapes that are too complex to describe explicitly can be
distinguished (e.g., coins showing a face or a symbol, floral vs. geometric ornaments, etc.).
Using the shape map: The hierarchical snapping process
Figure X shows mock-up examples, e.g., the decomposition of a column acquired from the HerzJesu church (Graz) into three basic cylindrical shapes, a capital (symmetric) and a box-shape on top.
The shape map is applied (step 2), e.g., by first clicking on the main cylinder, which triggers the
“snapping” of the first shape detector. The other two (collinear) cylinders are searched for in its
vicinity, and once they also snap in, the cylinders are elongated until the column capital is reached.
The next shape detectors are the symmetry and box detectors, which are supposed to look in a
certain search distance for mesh parts to snap to.
Figure X: Shape map mock-ups. Left: A capital detector might consist of a fitting box (top), a
symmetry plane (middle) in a segmentation box, and three fitting cylinders (bottom). Right: after
a rough initial alignment, the cylinders extend their height until the slippage condition is violated.
Planar regions can be found in many ways including slippage analysis, RANSAC plane fitting,
SVD, ICP.
This example illustrates that the snapping order is important. A good shape map can be applied by
using only a few mouse clicks to “anchor” some strategically chosen places of the model, thereby
spanning a reference frame for other detectors revealing more of the detail structure of the model.
The information flow is not only top-down, however, since information from later stages in the
snapping process can also be relevant for improving the accuracy of detectors that snapped in
earlier. It will also be necessary to realize a backtracking strategy to deal with the situation that
required detail cannot be found. Shape maps will internally also be represented as GML code. Note
that in this case, GML is used not for describing parametric shapes, but shape maps, i.e.,
hierarchically structured parametric shape detectors. In either case, however, the expressiveness of a
full programming language is required.
Shape Search
As the amount of 3D data becomes more common, and also outside the field of CR is rapidly
expanding (e.g. through portals such as Google Warehouse) the need for effective data-mining
(Funkhouser, 2004; Toldo, 2009; Ovsjanikov, 2011; Knopp, 2010) is rising too. In this project, we
will therefore also consider the task of improving shape retrieval with the help with the help of user
guided, and fully-automatic, unsupervised, generic and fast verification and expansion.
This work draws heavily from a variety of existing research in the related areas of content based 3D
retrieval and image-based search. Ideally the numerical description of shapes and the comparison
metric would be optimized for perfect retrieval. However, it is difficult to quantify semantic
concepts or to find the appropriate mathematical representation. Therefore, despite the large amount
of work done on finding generic detectors and descriptors suited for visual tasks (Lowe, 2004;
Proposal Part B: page 8
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
Willems, 2008), significant effort was also invested in devising better features (Akbar, 2007;
Papadakis, 2008; Philbin, 2010), attributes (Philbin, 2008; Ferrari, 2007), distances (Elad, 2002;
Akbar, 2007; JSegou, 2007) and projections. A major drawback of these approaches is the
requirement of user supervision, inability to generalize (in case of typical user-driven approaches)
and delay from incorporating relevance feedback into online learning. Generic off-the-shelf 3D
features (Willems, 2008) in BoW (bag of words) based methods (Toldo, 2009; Ovsjanikov, 2011;
Knopp, 2010) have been shown to be robust to noise, deformation, orientation etc. and we exploit
these in our work for the retrieval task. Interestingly, though BoW approaches have been used in
shape search, relevance feedback has mostly relied on vanilla features (Elad, 2002; Papadakis,
2008; Akbar, 2007).
Computational time and accuracy often have to be played off against each other, when examining
the semantic relevance of search. Regardless of the chosen representation, most methods end up
improving search relevance in a somewhat cascaded or iterative manner. For example, two (or
more) representations (feature or distance wise) for an object may be constructed: one in which
search can be performed fast and another in which accuracy can be improved. Improvement of
accuracy and performance is still to be achieved in a variety of ways:



incorporating user feedback (most popular in 3D (Minka, 1996; Elad, 2002, Papadakis,
2008; Hoi, 2011),
structural consistency (as in 2D (Philbin, 2008, Mikulik, 2010; Chum, 2011)
a variety of heuristics e.g. pseudo-relevance feedback, multiple queries etc. (see (Lou, 2003,
Bang, 2002; Papadakis, 2008).
Application use case: Plaster moulds for ornaments of the Regency period in Brighton
A paradigmatic example for the necessity of managing collections of similar shapes – in that case
even self-similar shapes – is the excellent collection of historical plaster moulds hosted by partner
BTCT.
NICK/PHIL: Please describe use case
Fig. X: Wooden moulds for plaster ornaments of the Regency period in Brighton (19th century).
There are over 700 of these original moulds, which are still in use today. However, subtle shape
Proposal Part B: page 9
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
differences make it quite difficult to find the right mould when carrying out reconstruction work.
1.2
Progress beyond the state-of-the-art
Shape Search
In this project several improvements user feedback, structural consistency, a variety of heuristics
e.g. pseudo-relevance feedback, multiple queries etc. The goal is to create a system that is able to
learn shapes unsupervised. Whenever certain feature correspondences have been found between
objects; an iterative scheme would try to combine different instances that corroborate one another.
Finally the object definition is not described anymore by its individual instances, but by its
expanded version, or simply stated its average (although that does not cover the real context)
In the quest for such a scheme, a 2-pass approach is planned based on query verification and
expansion. First, the results of a vanilla search scheme based on vanilla BoW based retrieval are
weakly verified according to the spatial configuration of the query object. This is followed by a
second pass which retrieves shapes similar to the verified set of results to create an expanded list.
While the original vanilla BoW search is fast, giving the method many possible candidates, the
more expensive verification and expansion searches in an alternative space, lead to an increase in
accuracy of results. The query expansion was found as an optimal retrieval method when searching
is defined as a classification task (Efron, 2008).
BoW representations succeed in representing a shape in terms of feature occurrence and being
orientation and noise invariant but fail to capture more layout information. An improvement over
vanilla search, was the analysis of word co-occurrences (Sivic, 2004; JSegou, 2009; JSegou, to
appear) and more recently, structural constraints (Chum, 2011; Arandjelovic, 2011). Actual shape
has more complex representation: e.g. statistical models, articulated models (among other graphs),
templates etc. Fitting such models to a given shape can involve working our parameterisation and
correspondences. An exhaustive and comprehensive fitting procedure can be expensive, rendering
this useless for the purpose of shape verification in search. Hence, we extend schemes that leverage
feature co-occurrence to take into account their mutual spatial layout. Though a very loose
approximation of more complicated shape models, this scheme of weak structural verification
proves very effective in improving search performance.
In the above scheme it is more or less assumed that the objects are defined by their overall shape.
This is in general also the main assumption in literature. However, shapes can be related to one
another as well because of correspondences that are found locally; i.e. on parts of the object. For
instance, a collection of statues or reliefs may correspond because they all contain faces; however,
the context in which they are presented may be completely different.
Proposal Part B: page 10
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
Therefore additional methodologies will be created to indicate areas of interest, which can then be
used to fine-tune the properties of these subshape, i.e. descriptors. The query software will enable to
look and track down objects in the database that show such partial features.
In addition, apart from being able to perform queries on weak(er) similarity, the approach will also
assist in the actual recognition of partial shape instances.
The example below shows an existing case in the transcription of cuneiform tablets. It appears that
many of these tablets appear in “bulks” and show identical signature configurations. These
configurations (indicated in color) serve as a means for classification. Whereas at this stage each
cunefirm has to be transcribed individually since there is no a priori knowledge, local shape search
would allow to perform an initial classification aiding tremendously in translating these tablets
more efficiently
<<<<
.
Proposal Part B: page 11
FP7-ICT-2011-9
18/01/12 v1
1.3
STREP proposal
SHAPEATLAS
S/T methodology and associated work plan
1.3.1
Work package list
Work
package
No
Work package title
Type of
activity1
Lead
partic
no.2
Lead
partic.
short
name
Personmonths
Management
MGT
1
FhA
36
2
Acquisition
RTD
4
KUL
45
3
Shape Map Definition
RTD
3
TUG
33
4
Shape Indexing
RTD
4
KUL
45
5
Shape Query
RTD
3
TUG
33
6
Integration and Evaluation
RTD
2
UOB
27
7
Dissemination
DEM
1
FhA
6
1.3.2
End
month
3
1
TOTAL
Start
month4
4
1
36
225
List of deliverables
Del. no. Deliverable name
WP no.
Nature6
5
Dissemi
-nation
level
7
Delivery
date8
(proj.
month)
1
Please indicate one activity (main or only activity) per work package:
RTD = Research and technological development; DEM = Demonstration; MGT = Management of the
consortium
2
Number of the participant leading the work in this work package.
3
The total number of person-months allocated to each work package.
4
Measured in months from the project start date (month 1).
5
Deliverable numbers in order of delivery dates. Please use the numbering convention <WP number>.<number
of deliverable within that WP>. For example, deliverable 4.2 would be the second deliverable from work package 4.
6
Please indicate the nature of the deliverable using one of the following codes:
R = Report, P = Prototype, D = Demonstrator, O = Other
7
Please indicate the dissemination level using one of the following codes:
PU = Public
PP = Restricted to other programme participants (including the Commission Services).
RE = Restricted to a group specified by the consortium (including the Commission Services).
CO = Confidential, only for members of the consortium (including the Commission Services).
8
Measured in months from the project start date (month 1).
Proposal Part B: page 12
FP7-ICT-2011-9
18/01/12 v1
1.3.3
STREP proposal
SHAPEATLAS
List of milestones
Milestone
number
Milestone
name
Work package(s)
involved
Expected date 9
9
Means of
verification10
Measured in months from the project start date (month 1).
Show how you will confirm that the milestone has been attained. Refer to indicators if appropriate. For example: a
laboratory prototype completed and running flawlessly; software released and validated by a user group; field survey
complete and data quality validated.
10
Proposal Part B: page 13
FP7-ICT-2011-9
18/01/12 v1
1.3.4
STREP proposal
SHAPEATLAS
Work package description
Work package number
Work package title
Activity type11
Participant number
Participant short name
Person-months
per
participant
1
Start date or starting event: Month 1
Management
MGT
1
2
3
4
5
FhA
UOB
TUG
KUL
BTCT
18
18
Objectives
WP1 covers all activities related to project management and reporting to the commission. Its primary goal is
to ensure that research work is focused and carried out according to the overall project vision and work plan,
within the time and budget constraints and with efficient use of available resources.
Description of work (possibly broken down into tasks) and role of partners
The overall approach to management is described in Section 2.1 and implemented via the specific tasks
described below. WP1 will entail day-to-day central management activities, including establishment and
maintenance of tools for intra-project communication and collaboration, development and maintenance of
the project website and its content management, communication and collaboration infrastructure,
organization of project meetings, and the development and delivery of project reporting, both internally to
the project and externally to the Commission. Furthermore, WP1 will undertake the preparation and
implementation of quality plans, establish and execute respective monitoring processes, and establish and
implement risk management and contingency plans.
The project management and coordination will be performed by Fraunhofer Austria (with focus on technical
coordination) and University of Brighton (with focus on administrative management):
Task 1.1. Quality assurance & risk management (FhA, UBO)
This Task will formulate and outline the procedures for quality assurance and risk management to be
followed and maintained for the management of the project. The output of this Task is a Quality Assurance
& Risk Management Plan (QA&RM). This Task also covers a continuous assessment of project uncertainties
and challenges and proactive elaboration and implementation of contingency measures, as well as corrective
actions if and where needed. This plan will incorporate, elaborate on and extend potential risks as identified
at the proposal stage.
Task 1.2. Knowledge management and IPR resolution (FhA, UBO)
The consortium’s approach to intellectual property rights (IPR) issues and the management of knowledge
produced during the project is presented in Section 3. Before the project starts, the partners will draft a
Consortium Agreement for the management of Knowledge Produced (as an Annex to the Grant Agreement),
which identifies their pre-existing know-how which they may grant access rights to the consortium (listing
also relevant pre-existing know-how excluded from the obligation to grant access rights). During the project,
this task will undertake the implementation and elaboration of this agreement. In specific, in order to clarify
any controversies or disputes, a detailed IPR management plan will be developed during the first months of
the project and will be updated throughout the project.
Task 1.3. Communication & Collaboration (FhA)
This task will establish the appropriate communication and coordination infrastructure and ensure the smooth
communication among consortium partners, and with the EC. Specific activities include:
 establish and maintain mailing lists and an on-line real-time communication platform (e.g. skype or
other teleconferencing tool);
 establish and maintain the internal project web-based collaboration platform and document
11
Please indicate one activity (main or only activity) per work package:
RTD = Research and technological development; DEM = Demonstration; MGT = Management of the consortium.
Proposal Part B: page 14
FP7-ICT-2011-9
18/01/12 v1



STREP proposal
SHAPEATLAS
repository;
maintain procedures for monitoring internal communication and collaboration (as will be detailed in
the QA&RM Plan);
ensure and monitor smooth communication within the consortium, with EC and with the External
Advisory Board;
organize and manage project meetings (including detailed agenda and minutes for each meeting).
Task 1.4. Administrative & financial management (UBO)
This task will initiate, set and implement the consortium contract, including the consortium agreement, the
overall administrative management processes and routines. This task will also take care of all the financial
aspects and routines related to the project. In particular specific actions are:
 establish and monitor project internal reporting schedules and procedures regarding administrative
and financial issues, including 3-month internal progress reports;
 prepare the necessary progress reports and financial records according to EC guidelines;
 support to partners in completing their contributions to the Periodic Reports which include the Cost
Claims and any required Certificate on the Financial Statements (CFS)
 obtain Audit Certificates (CFS) from partners (where applicable).
Deliverables (brief description) and month of delivery
M.1.1. Quality Assurance & Risk Management Plan (Month 1)
The QA&RM plan will define in detail all procedures for quality assurance in project communication,
collaboration and deliverables.
M.1.2. IPR management plan (Month 1)
The initial version of IPR management plan (M01) will detail the plan and specific procedures needed to
implement the Consortium Agreement (Annex to the Grant Agreement) with respect to knowledge
management. Subsequent versions (M13, M25) will include detailed descriptions of project foreground
knowledge and (if needed) amendments to the Consortium Agreement.
M.1.3. Project collaboration & communication infrastructure (Month 1)
This deliverable entails the web-based collaboration & communication platform that will be used during the
project. Content and activities will be developed during the course of the project. An accompanying short
report will present a basic outline of the infrastructure and its use.
M.1.4. Project Progress and financial reports (Month 1)
Regular progress and financial reports as mandated by the Grant Agreement and the EC. These include
yearly and final project report, or any other report that may be requested by EC.
M.1.5. Updates to D.1.1. – D.1.4. (Month 13)
Continuously performed updates and modifications are collected and reported.
M.1.6. Updates to D.1.1. – D.1.4. (Month 25)
Continuously performed updates and modifications are collected and reported.
M.1.7. Updates to D.1.1. – D.1.4. (Month 37)
Continuously performed updates and modifications are collected and reported.
D.1.8. Final Report (Month 38)
The final, financial report documents the project’s scientific, economic and social success.
Work package number
Work package title
Activity type
Participant number
Participant short name
2
Start date or starting event:
Acquisition
RTD
1
2
3
4
5
FhA
UOB
TUG
KUL
BTCT
Proposal Part B: page 15
FP7-ICT-2011-9
18/01/12 v1
Person-months
participant
STREP proposal
SHAPEATLAS
per
12
27
6
Objectives
The Brunswick Town Charitable Trust defines the use-case scenario, in which they provide 3D objects.
These objects are photographed by UOB and processed by KUL. The reconstruction process will need some
adjustments and new features (multi sequence processing, etc.), which will be added and implemented within
this work package.
Description of work (possibly broken down into tasks) and role of partners
Task 2.1 Improvement of photogrammetry acquisition technique (KUL)
KUL will provide the necessary tools for 3D acquisition.
 The acquisition is chosen to be image based, in order to provide a flexible non-expert approach for
the 3D digitization. Proper procedures will be defined to create image sequences, which will be
processed by a cloud based reconstruction service called Arc3D.µ
 Special effort will be put into allowing multi-view sequences. In the current state of the art it is quite
common to enforce single continuous sequences, however, for many type of models different paths
are necessary in order to cover the whole object. It is further investigated whether the intermediate
knowledge on the shape can also assist in defining these paths.
 An assessment will be carried out on additional complementary techniques that involve additional
structured light and photometric stereo, in order to verify whether they enhance the acquisition
pipeline for the given dataset or business case.
 Since the acquisition pipeline is fully self-contained and is not dependent on third parties, the
scanning can be accompanied by immediate shape analysis. Therefore the immediate output is only
restricted to the 3D model, but also shape related characteristics will be integrated and can be fed
back into the system.
Task 2.2 Define and develop user case scenarios (UOB, BTCT)
The experiments will be designed and conducted by groups other than those responsible for the development
of the acquisition tools. The intention is to acquire enough 3D content using the improved photogrammetry
tool Arc3D in order to test the shape map core technology. The result of each experiment will be documented
in a 3D repository along with the metadata regarding its capture. It is expected that the data for this scenarios
will include ornaments of heritage buildings as well as architectural elements.
Task 2.3 Documentation of domain knowledge on shape classification (UOB, BTCT)
This task will develop the documentation on the domain knowledge of the user case scenarios, in particular
regarding how shape is defined according to the experts in the area. This knowledge will be documented and
will serve as input for user requirements to support the development work packages (WP3, WP4, WP5,
WP6).
Deliverables (brief description) and month of delivery
D.2.1. User case scenarios definition and protocols for implementation (Month 6)
D.2.2. Domain knowledge and user requirements based on user case scenarios (Month 9)
D.2.3. Alpha version of photogrammetry techniques (Month 12)
D.2.4. Document acquired content (Month 24)
D.2.5. Beta version of photogrammetry techniques (Month 24)
Proposal Part B: page 16
FP7-ICT-2011-9
18/01/12 v1
Work package number
Work package title
Activity type
Participant number
Participant short name
Person-months
per
participant
STREP proposal
SHAPEATLAS
3
Start date or starting event:
Shape Map Definition
RTD
1
2
3
4
5
FhA
UOB
TUG
KUL
BTCT
12
27
Objectives
TU Graz realizes the concept of local shape maps; i.e. a user-defined regions of interest, which have special
features. These regions and features will have a special weighting in the feature-based clustering within the
shape indexing (see WP 4). To describe a shape template (consisting of these shape maps) these local shape
maps have to be combined using a graph structure (called atlas). UOB will define these graph structures.
Description of work (possibly broken down into tasks) and role of partners
A shape map is a user-defined model for 3D shape measurement and classification. It consists of elementary
parametric shape detectors that are combined in a hierarchical fashion to obtain a user-defined pattern
matching template for a complex shape. The objectives of this work package are to develop the shape map
core technology, and to create an easy-to-use interactive 3D software application for the creation of shape
maps. This application shall enable users with a non-technical background (e.g., from the Cultural Heritage
sector) to explicitly describe, for the shapes of a collection, which particular properties of a shape constitute
its class membership, which shape parameters are to be measured, and where on the shape they are
measured.
In particular, the tasks in WP 3 are:
 Primitive shape detectors (see Section 1.1): Min/max probes, segmentation volumes
(box/sphere/cylinder), fitting volumes (box/sphere/cylinder), crease/edge/corner detector, slippage
analysis, orientation plane, symmetry plane, principal axis, similarity search region.
 Shape map hierarchy (see Section 1.1): The multi-level snapping proceeds along the hierarchy where
sub-detectors that are defined in the reference frame of their parent detector. Possible positions and
orientations of the child can be prescribed.
 Shape map software: It offers a set of tools to create, place, and configure primitive detectors, and to
refine the hierarchy by specifying search constraints for sub-detectors. This tool is used both for
defining (once) and for instantiating (many times) a shape map.
In the tools the emphasis is on the efficient computations that allow interactive speed, which is for some
detectors are very challenging task.
Deliverables (brief description) and month of delivery
Work package number
Work package title
Activity type
Participant number
4
Start date or starting event:
Shape Indexing
RTD
1
2
3
4
5
Proposal Part B: page 17
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
FhA
Participant short name
Person-months
per 18
participant
UOB
TUG
KUL
17
BTCT
Objectives
feature vector-based clustering and local feature detection (KUL), graph matching for shape maps (FhA)
Description of work (possibly broken down into tasks) and role of partners
Task 4.1 (KUL)The goal of this task is to improve 3D shape retrieval, with help of query verfication and
expansion in an automatic, unsupervised and class-generic setup. In addition to shape description based
on established features and visual words, we employ orientation-invariant structural similarity to improve
search relevance. We plan a cascade method consisting of two steps:


After the first inexpensive step of retrieval using a standard BoW (Bag-of-Words) based
approach, a simple, fast but effective weak spatial layout verification step is used to prune
the initial search result.
A new BoW query is constructed and issued for an expanded result. We perform
comprehensive evaluation and check whether or not improved performance on retrieval and
classification exists on the datasets (especially for limited training data)
We expect substantial improvement in the description of the shape attributes and therefore as well in the
retrieval performance of the dataset.
Task 4.2 (KUL)
This task will more particularly be dedicated to the proper description of partial surface patches that serve
special interest based on user feedback. The similarity or correspondence measurements on local features
will be finetuned to also assist in recognition of shape features rather than query of weak similarity.
Deliverables (brief description) and month of delivery
D4.1 Intermediate status on shape indexing
D4.2 Final status on shape indexing
Work package number
Work package title
Activity type
Participant number
Participant short name
Person-months
per
participant
5
Start date or starting event:
Shape Query
RTD
1
2
3
4
5
FhA
UOB
TUG
KUL
BTCT
6
21
10
Proposal Part B: page 18
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
Objectives
Interactive online application for searching and browsing the shape database showing a dynamic live result
set, combining shape-based search and parametric search. Parameter selection either form-based or by
automatic parameter clustering through relevance feedback. Refined shape maps are offered for searching
sub-classes in order to restrict and specialize the scope of the search.
Description of work (possibly broken down into tasks) and role of partners
This WP is concerned with the end-user frontend for querying the shape database. The challenge is to
combine shape-based search with parameter-based search in a seamless way. The interface will support
query by example as well as form-based search, providing the shape parameter set of the chosen shape class
in a refineable way to allow also searching for more specialized sub-classes (from hat to helmet); but also
generalization is possible, going from a specific shape to the super class (from helmet to hat).
The specific tasks are:
 Relevance feedback: The user can select the most relevant of the currently offered search results.
The result set is modified according to (i) shape similarity and (ii) parameter similarity, using
automatic parameter clustering to determine the most probable search direction.
 Proper similarity measures will be defined based on the use of a symmetric Hausdorf_distances
between the two shapes. A weak shape model approach will be introduced to allow small
dissimilarities and / or geometrical distortions to cover possible variability in shapes.Shape
map editing: Every object shown as search result has one or more associated shape maps. The user
can select one of them and also alter the instance parameters, and the result set is updated
accordingly (select thin amphora, increase girth parameter, result set: thick amphorae)
 Parameter inspection: The user can select any search result and look at the instance parameters of the
shape map; select one parameter; and re-order the search result with respect to this parameter.
 Specialization/generalization: Since the shapes in a result set are similar, they have similar
specializations and generalizations. The respective shape maps are shown to direct the search.
 Representative instance SVEN
As in WP3, interactivity is the key for maintaining a fluent and targeted search. This will require developing
optimization strategies for computational efficiency.
Deliverables (brief description) and month of delivery
D.5.1.
Work package number
Work package title
Activity type
Participant number
Participant short name
Person-months
per
participant
6
Start date or starting event:
Integration and Evaluation
RTD
1
2
3
4
5
FhA
UOB
TUG
KUL
BTCT
12
6
9
Objectives
This work package will create an integrated platform with the shape map core technology, the shape
indexing and query mechanisms, which allow a user to explore a collection. Furthermore, it will undertake
Proposal Part B: page 19
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
an evaluation of the tools developed as effective solutions to the problems addressed by the project.
Description of work (possibly broken down into tasks) and role of partners
Task 6.1 Integrated platform for search and visualisation (UOB)
This task will develop an integrated platform integrating the different technologies developed within the
work packages into a tool which can be deployed in the cultural heritage sector.
Task 6.2 Evaluation of technology (UOB, BTCT)
This task will test different aspects of the tools for 3D artefacts in controlled experiments designed to
evaluate both individual tools and the working practices their use implies. The results will be a systematic
evaluation used to inform improvements to the tools and to understand the requirements for deployment in
the field. These experiments will be developed using different case scenarios provided by the heritage
organisations.
Deliverables (brief description) and month of delivery
D.6.1. Alpha version of integrated platform (Month 24)
D.6.2. Testing protocols (Month 30)
D.6.3. Feedback from testing (Month 36)
Work package number
Work package title
Activity type
Participant number
Participant short name
Person-months
per
participant
7
Start date or starting event:
Dissemination
DEM
1
2
3
4
5
FhA
UOB
TUG
KUL
BTCT
3
15
3
Objectives
 Disseminate the project activities and results supported by different mechanisms, including publications
and events.
 Create and moderate a website supported by the project partners.
 Provide quality publications for the EU digital cultural heritage community.
 Explore the potential for exploitation.
Description of work (possibly broken down into tasks) and role of partners
Task 7.1 Website (UOB, FhA)
The website aims to support those working in digital cultural heritage by providing several avenues for
professionals to connect and access cutting edge information, tools produced by the project and events. The
website will be updated constantly by the project partners and will have different levels of security and
access depending on user status.
Task 7.2 Publishing (UOB, FhA)
The project will provide a biannual newsletter that will serve to inform, engage and stimulate professionals
in digital cultural heritage. The newsletters will focus on summaries of current project news, issues, research
Proposal Part B: page 20
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
and events.
Task 7.3 Dissemination and demonstration events (UOB, FhA)
Biannual events will be organised at existing and ongoing conferences such as CAA, VAST, VSMM. These
events aim to:
 Connect with relevant professionals,
 Demonstrate and disseminate the project results,
 Encourage those working in CH to deploy the technologies developed by the project
This will also include the production of dissemination material to support these events, such as the
production of printed material (e.g. publications, leaflets), video, and participatory fees.
Task 7.4 Exploitation (UOB, BTCT)
This task will map out the potential market and opportunities for exploitation of the project outputs. It will
adopt a layered approach, considering the specific opportunities associated with the use-case scenario
dataset, the broader potential for application to a range of reconstructive CH scenarios, and the general
exploitability of the tools and techniques developed, individually or in combination. It will use the
dissemination activities and evaluation studies to engage with potential clients, identify exploitation
scenarios, map out strategies and assess risks. It will produce an Exploitation report summarising its findings
and making recommendations for sustainable exploitation after the project has ended.
Deliverables (brief description) and month of delivery
D.7.1. Website (Month 1, 30)
D.7.2. Report of website (Month 7, 14, 21, 29)
D.7.3. Results on networking activities (Month 12, 24)
D.7.4. Exploitation report (Month 30)
Proposal Part B: page 21
FP7-ICT-2011-9
18/01/12 v1
1.3.5
STREP proposal
SHAPEATLAS
Summary of effort
Partic.
no.
1
2
3
4
5
Total
Partic.
short
name
FhA
UOB
TUG
KUL
BTCT
WP1
WP2
WP3
18
18
12
12
27
36
27
6
45
WP4
WP5
WP6
6
21
12
6
27
9
27
18
WP7
3
15
27
39
45
Proposal Part B: page 22
3
21
Total
person
months
39
75
54
54
18
240
FP7-ICT-2011-9
18/01/12 v1
2
2.1
STREP proposal
SHAPEATLAS
Implementation
Management structure and procedures
The project management will be planned to guarantee delivery of the results on time and to manage project
complexity. General project organization covers both technical and administrative issues.
The consortium management of SHAPEATLAS intends to reach the following objectives:
 To manage communication among consortium partners.
 To manage technical, financial, legal and administrative activity of the consortium.
 To assure that the end product conforms to the planned requirements and product description.
 To preserve efficient communication with the European Commission.
 To satisfy compliance with EC standards and procedures for project management.
The project management plan includes the following activities: The administrative and technical
management will carry out the project management. It will control decisions about communication among
consortium members and project implementation. The implementation will also comprehend coordination
and support for reports and financial management. The coordination shall also be used for efficient
exploitation of the project outcomes.
2.1.1 Management roles
The project management consists of a technical project manager and an administrative project
manager. While the administrative project manager manages project resources, advertises the
project and handles the project dissemination, the technical project manager supervises consortium
performance, ensures the project result’s quality and represents the project and consortium –
especially the technical project manager shall communicate with external organizations.
Work Package leaders are responsible for reporting and follow up of deliverables and milestones of
each particular work package. They also efficiently coordinate tasks in particular work packages.
Furthermore, they will initiate and participate to the scientific/ technical meetings necessary for
work progress and report minutes. Each WP leader is responsible of ensuring the accomplishment
of the scientific and technical objectives of the WP, by assessing the quality of the outputs of the
performed work and solving local conflicts involving the tasks execution. In case of failures or
major issues that affect the completion of the work foreseen, the WP leader must refer to the
technical project manager.
2.1.2 Decision-making bodies and mechanisms
The project board is the advanced representation and management body of the project, led by both
project managers (administrative and technical). It will be responsible for:
 Resolving and monitoring project's technical progress, taking care among work packages,
implementation and directs the execution of the Implementation plan, and preparation the
reports.
 All formal decisions, strategic guidance, communications with European Commission, plans
for promotion of project outcomes, etc.
 Applying procedures for managing intellectual property rights and innovative outcomes of
the project.
 The implementation of a contingency plan based on the assessment of indicators and
identified risks will be guaranteed by the project board members.
The project board will meet at least once every 4 months. Each consortium member will have one
representative in the project board.
2.1.3 Reporting mechanisms
Each consortium partner will create a management report every three months about the progress of
their work. These reports will be delivered to the technical project manager, who will combine
them into a unified report that will be delivered to the European Commission twice per year. The
Proposal Part B: page 23
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
technical project manager will revise the reports to guarantee the completeness and consistency.
The reports contain all activities accomplished by the consortium members in the corresponding
period: research and development activities, deliverables, meetings, etc.
2.1.4 Meetings, conferencing and communication tools
The consortium members will have regular meetings and discussions using various forms of
modern communication – especially Voice-over-IP techniques (Skype), teleconference systems, etc.
as a regular alternative to physical meetings in order to optimise the use of budget (reduce travel
costs and travelling time) and to reduce carbon footprint. The project will be handled wisely to
avoid high travelling cost and will be aimed to achieve maximum information exchange between
the consortium members. The general consortium meeting will take place periodically. If necessary,
intermediate meeting among smaller teams will be held. The collaboration between the consortium
members will be arranged to keep track of the project activity, obtain knowledge and deliverables.
The consortium members will have opportunity to transfer, publish, share information and comment
the common agenda.
2.1.5 Knowledge management and intellectual property rights
The intellectual property rights will be managed with appropriate procedures according to the
consortium agreement. The existing intellectual property rights and their knowledge will be
determined and guaranteed with agreed procedures before the start of the project. Any intellectual
right originated in the project will be determined and regulated in the consortium agreement that
will be prepared and considered by all consortium members.
Special activity will be taken to identify background knowledge and technologies of the consortium
partners before the start of the project.
2.1.6 Quality assurance and control
The technical project manager will be responsible for the quality assurance. The quality assurance
procedures will guarantee that the highest technical and scientific qualitative of deliverables and the
requirements defined in the product description in the planning process will be ensured. The
quantitative and qualitative indicators will be applied to evaluate the project performance towards
meeting its objectives. Each work package leader will be responsible to check periodically whether
the deliverables meet the quality standards. Any comments and suggestions from the consortium
partners how to improve the quality of deliverables will be considered. Similarly, the administrative
project manager monitors and measures costs and schedule performance.
Before submitting any document to the European Commission, a peer review process within the
project board members will be performed.
2.1.7 Conflict management
The project consortium will consist of diverse members. Because potential conflicts are possible,
some basic guidelines for its management will be prepared and lead by the project management.
Conflict management will manage conflicts between:
 Individual members goals and project goals as well as
 the potential disagreement according to the project's schedules and priorities.
The following plan will be prepared:
Establish Responsibilities. The technical and administrative project managers will establish
responsibility by taking charge of resolving and managing the conflict. All consortium members
have the responsibility to report any determined issues before they become conflicts.
Establish Conflict Management Strategy. The technical and administrative project managers follow
a strategy to:
 detect and solve issues before they become severe conflicts,
 establish a trusted environment for the consortium members to exchange ideas, and
Proposal Part B: page 24
FP7-ICT-2011-9
18/01/12 v1

STREP proposal
SHAPEATLAS
engage the consortium members to express their minds.
If this strategy fails, the project board has to solve a conflict by majority vote.
2.2
Individual participants
2.2.1 Fraunhofer Austria Research GmbH
Organization profile
Fraunhofer Austria (FhA) is a fully‐owned subsidiary of the German Fraunhofer‐Gesellschaft. Its
research areas include Visual Computing & Digital Libraries, located in Graz and Production &
Logistics in Vienna. The business area in Graz belongs to an excellence cluster on computer
graphics and computer vision with connections to Technische Universität Graz (TUG). Research
undertaken by Fraunhofer is directly aimed at promoting industrial performance, which
distinguishes Fraunhofer from other large research institutions involved in pure or basic research.
This focus applies equally to contract research for industry or government, as well as to advanced
strategic research.
Key competences and role in the project
The main focus of Fraunhofer Austria at Graz includes three research aspects: digital society, visual
decision support and virtual engineering. In the context of SHAPEATLAS these topics are of
special interest. Fraunhofer Austria develops technologies in order to capture and expand
knowledge. Fraunhofer Austria Visual Computing, as well as the Institute of Computer Graphics
and Knowledge Visualization of TUG are led by Prof. Dr. techn. Dieter W. Fellner. Both research
groups have 20 researchers in total and share a common infrastructure. They have gained expertise
on semantic modeling, immersive visualization, and physics‐based simulation in various national
and international projects.
Fraunhofer Austria will coordinate the project and manage its technical aspects. Having experience
in shape description and semantic enrichment, FhA will also contribute to WP4 “Shape Indexing”.
Furthermore, FhA will use its networking activities to promote SHAPEATLAS and its results.
CVs of key personnel
Prof. Dr. techn. Dieter W. Fellner is head of the Institute of Computer Graphics and Knowledge
Visualization at TUG and head of Fraunhofer Austria. He is also head of Fraunhofer Institut f.
graphische Datenverarbeitung (FhG‐IGD), the world’s leading institute for applied visual
computing, with applications in medicine, automotive industries, urban management (3D‐GIS),
virtual and augmented reality, and many more. His main research areas are computer graphics and
digital libraries. As principal investigator he has led a strategic initiative on Digital Libraries since
fall 1997, funded by the German Research Foundation (DFG) over a period of seven years. He has
written a widely used book on Computer Graphics (1988, 2nd ed. in 1992) and co‐authored a book
with A. Endres on Digital Libraries (2000). He has regularly served on editorial boards of leading
journals in computer graphics and digital libraries.
Dr. rer. nat. Eva Eggeling holds a PhD in Applied Mathematics from University of Cologne
(2002). She has developed numerical simulations in the application fields of meteorological data
assimilation and grain growth simulation for polycrystals. She has experience in modelling and
simulation in virtual reality, visualization of scientific data and differential and parametric
optimization. After receiving her PhD she was responsible for scientific coordination at the
Fraunhofer Institute for Algorithms and Scientific Computing. In 2006 she spent 3.5 years at
Carnegie Mellon University (Pittsburgh, USA). There she worked in an interdisciplinary team in the
area of material science simulation. Eva Eggeling is head of the business area Visual Computing of
Fraunhofer Austria.
Proposal Part B: page 25
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
Dr. techn. Torsten Ullrich has a PhD in computer science from Graz University of Technology
(2011) and an MSc in mathematics from the Karlsruhe Institute of Technology. His main research
areas are generative modelling, geometric reconstruction and optimization. He combined these
topics in his PhD-thesis “Reconstructive Geometry” to an inverse optimization approach, which
uses generative models for reconstruction purposes.
2.2.2 University of Brighton
Organization profile
The University of Brighton is a public teaching and research university, with approximately 2100
staff and 19000 students spread across 4 campuses along the south coast of England. It is one of the
leading ‘post-1992’ universities for research in the UK, with a strong research culture in computer
science and cultural informatics. The Cultural Informatics Research Group was coordinator of both
the EPOCH Network of Excellence in FP6 and the 3D-COFORM IP in FP7, and a partner in
CHIRON, a Marie Curie Research Network in Cultural Heritage Informatics. The Natural Language
Technology Group has a track record of high quality research in natural language processing
stretching back to the early 1990’s. The group’s research interests include text analysis, text
generation and linguistic knowledge representation, encompassing traditional formal language
methods (grammars etc.), and modern statistical approaches to language processing.
Key competences and role in the project
The technical focus of the Cultural Informatics group’s work is in bespoke 3D modelling and
rendering systems, led by Dr Karina Rodriguez-Echavarria. The group also led the development of
the EPOCH Network of Expertise Centres and EPOCH’s work in the area of assessing socioeconomic impact, which was reported in the EPOCH final review report as a highlight making
“ground-breaking progress in developing innovative methods and theory in the economics of the
cultural heritage”. This work has continued to be developed by Dr Jaime Kaminski under the 3DCOFORM project, including the establishment and hosting of VCC-3D, the Virtual Competence
Centre for 3D Technology, a not-for-profit enterprise venture to promote and support the
exploitation of 3D technologies. The Natural Language Technology Group’s expertise in language
processing has increasingly been applied to the processing of metadata and textual descriptions
associated with Cultural Collections (led by Dr Roger Evans).The group has also started exploring
the application of the same formal language techniques to problem of shape definitions, particular
in the context of procedural shape modelling systems.
CVs of key personnel
Professor David Arnold is Director of Research Initiatives for the University and Dean of the
Brighton Doctoral College. He is also Professor of Computing Science and was coordinator of both
EPOCH and 3D-COFORM. He has a 40-year career of research in the design of interactive
computer graphics systems and their applications in architecture, engineering, cartography,
scientific visualisation, health and most recently cultural heritage. Whilst at UEA, Norwich he led
that University's contributions to a number of EU projects (including CHARISMATIC). He was the
founding Editor-in-Chief of the ACM Journal on Computing and Cultural Heritage and has been
involved at a senior level in ACM, Eurographics, CEPIS and BCS. He has been one of the leaders
of the steering group that has established the VAST series of conferences over the past 12 years.
Dr Roger Evans is a Reader in Computer Science in the School of Computing, Engineering and
Mathematics, and research team leader in NLTG. He has over 22 years post-doctoral research and
management experience, spanning text analysis and generation, lexical representation and
architectures for natural language processing. He is a former SERC Advanced Fellow, a member of
the EPSRC College, a senior visiting research fellow at the University of Sussex, and former chair
of ACL-SIGGEN, the international SIG for Natural Language Generation. He has been PI on 8
EPSRC/SERC grants, coordinator of a 9-site EU (INTAS) project, site coordinator for 4 other EU
Proposal Part B: page 26
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
and international projects. He has been involved in the 3D-COFORM project contributing to the
development of metadata extraction and manipulation tools.
Dr Karina Rodriguez-Echavarria was awarded a PhD for her work on "Knowledge based
systems to support collaborative product development". She obtained her MA in Histories and
Culture at the University of Brighton in 2008. Karina works in the 3D-COFORM integrating project
in the area of 3D visualisation and the deployment of these technologies in cultural heritage
organisations. She previously participated in the European Network of Excellence EPOCH. Her
research interests include the documentation and visualisation of heritage collections, information
and knowledge management of 3D artefacts, and the practical aspects of deployment in the heritage
sector.
Dr Jaime Kaminski is an interdisciplinary researcher who began his career as an archaeologist. He
has a Ph.D. from the University of Reading which considered the archaeological evidence for
ancient environmental impact. He left archaeology in 1996 to become an analyst in a technology
research company, where he edited over 50 consultancy reports on technology issues and ICT
implementation in large enterprises. As a technology analyst he has undertaken freelance research
and consultancy projects for government, and other public and private organisations internationally.
Since 2004 Jaime has worked at the University of Brighton's Business School. Initially his research
focused on the socio-economic impact of cultural heritage sites, with specific reference the impact
of ICT on those sites as part of the European Commission-funded EPOCH Network of Excellence
and he leads the University’s contribution to the V-MUST NoE. Since 2008 he has worked on the
EC's 3D-COFORM project where he is a work-package leader. In this role he brings together
research on socio-economic impact, business modelling and sustainability as applied to the pipeline
of 3D data acquisition and visualisation. As part of this research activity he has co-developed
numerous impact assessment and strategy models for heritage. Furthermore, he has conducted a
great deal of research and consultancy in the field of social enterprise; in doing so becoming the
University's first 'Commercial Fellow' in Social Enterprise.
2.2.3 Technische Universität Graz
Organization profile
Graz Technical University was founded 1811 and is the second biggest technical university in
Austria with about 11.000 students in all fields of engineering. The Institute of ComputerGraphics
and KnowledgeVisualization (CGV) was founded in 2005 as part of the Faculty of Informatics. Its
mission and research focus is to link 3D data with semantic information in various fields, ranging
from virtual reality the over geometry processing and shape modelling to and digital libraries. CGV
has built its own Definitely Affordable Virtual Environment (DAVE), one of the largest CAVE
systems in Austria. The team of eleven PhD studentds and two post-doc researchers is led by Prof.
Dieter Fellner. Recent projects are German DFG funded (PROBADO on digital libraries), Austrian
FFG funded (Metadesigner on product mass customization, CITYFIT on procedural urban
reconstruction, AUTOVISTA on 3D surveillance), and EU funded (AGNES and V2me in ambient
assisted living, EPOCH and 3D-COFORM in Cultural Heritage), as well as industry projects
(SurfaceReconstruction project with Volkswagen AG). One of the core competencies of CGV,
which will be important in SHAPEATLAS, is procedural shape modelling on the basis of the
Generative Modeling Language. CGV is hosting the GML homepage under www.generativemodeling.org.
Key competences and role in the project
In SHAPEATLAS, partner TUG will contribute its long-standing expertise in procedural shape
modelling. The Generative Modeling Language (GML, see www.generative-modeling.org) is a
programming language for shape that was applied and extended in various research projects. Its
distinctive feature is that it greatly facilitates automatic code generation, which makes it ideal, e.g.,
for capturing the construction history of a complex 3D model during interactive shape modelling in
form of an executable script. So the end result of the modelling process is not just a mesh or a set of
Proposal Part B: page 27
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
NURBS patches, but the shape construction history itself. It can be re-evaluated to obtain a whole
collection of similar shapes. In SHAPEATLAS, however, the parametric infrastructure of GML will
be used for the inverse problem, shape analysis. Shape maps are composed of hierarchical shape
descriptors, which can suitably be described using GML’s operator paradigm.
CVs of key personnel
Prof. Dieter Fellner is full professor of computer science and the head of CGV at TU Graz, as well
as of GRIS at TU Darmstadt. Furthermore he is director of Fraunhofer IGD in Darmstadt, one of
Europe’s largest computer graphics groups with more than 100 full-time equivalent researchers.
Fellner has lead the Digital Library initiative V3D2 of the German Research Foundation (DFG)
which featured 20+ research projects over eight years. Fellner has written two books, numerous
research papers, was leader of the Eurographics Association, and is member of the editorial board
of various journals and conferences. He initiated and actively pursues the notion of 3D models as
generalized documents, which is the conceptual basis leading to SHAPEATLAS.
Dr. Sven Havemann is post-doc researcher at CGV, leading the generative modelling group. He
graduated in Bonn university, got his PhD (with distinction) from Braunschweig Technical
University in 2005 and his Habilitation from TU Graz in 2012. His research interests range from
industrial shape design and product mass customization over shape grammars and urban
reconstruction to virtual reality, advanced user interfaces, and shape repositories, typically in
connection with generative modelling. With more than ten years of experience in Cultural Heritage
projects (CHARISMATIC, EPOCH and 3D-COFORM) he will also contribute his knowledge of
practical problems in this sector, e.g., the management of large collections of similar shapes in a
repository of scanned shapes in a museum.
Martin Schröttner is currently finishing his master’s thesis at CGV under the supervision of Sven
Havemann. While his thesis is on hierarchical radiosity, i.e., a rendering method for global
illumination, Schröttner was also deeply involved in 3D-COFORM. He is responsible for
developing the MetadataGenerator tool for capturing process information of 3D-datasets in a
sustainable way using the CIDOC CRM standard, an entity-relationship model for semantic
networks.
2.2.4 Katholieke Universiteit Leuven
Organization profile
The team at the Katholieke Universiteit Leuven that will work on the project is part of the Center
for the Processing of Speech and Images (PSI), one of the units within the department of Electrical
Engineering (ESAT). The VISICS team (VISion for Industry, Communications, and Services) specialises in computer vision and its applications, in 3D acquisition and modelling, and on object
and object class recognition.
The team has received several prizes for its work, including a David Marr award, two TechArt
prizes, an EITC Technology Award, a Henry Ford prize and numeral best paper wards. It has
founded multiple spin-off companies: ICOS (chip inspection), Eyetronics (3D acquisition
technology for the games and movie industry mainly) and GeoAutomation (mobile mapping, i.e. 3D
measurements in cities from a moving van). It has been a partner in many European, Belgian and
Flemish projects.
Key competences and role in the project
The VISICS team has proed its expertise on 3D-digitisation and analysis for various applications
ranging from ranging from s in the cultural heritage sector. Their research activities address all
aspects of 3D-capture, 3D-processing, the semantics of shape, material properties, metadata and
Proposal Part B: page 28
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
provenance, integration with other sources (textual and other media); search, research and
dissemination to the public and professional alike.
The work lies at the basis of several spin-off companies, bringing 3D to various application
domains, varying from industrial inspection, medical application, special effects.
CVs of key personnel
Prof. Luc van Gool heads both this group and the BIWI computer vision group at ETHZ in
Switzerland. He is an electrical engineer (University of Leuven) and a full professor. He is editorin-chief of the journal Foundations & Trends in Computer Graphics and Vision, and a member of
several editorial boards and IPCs for major conferences. He has been involved as program chair and
area chair of the main vision conferences ICCV, CVPR, and ECCV several times. He has been
program chair of ICCV05, general chair of ICCV11 and will be general chair of ECCV14.
Dr. Marc Proesmans is currently responsible for project and innovation coordination at KUL. He
is an electrical engineer (University of Leuven) and did his PhD in early visual processing and 3D
reconstruction. The research results have spun off a company located in Los Angeles, specialized in
human scanning for movie / game VFX. He has over 20 years research, development and business
experience in the field of image processing and 3D scanning, and continues to perform and
supervise research activities on novel 3D acquisition techniques, with a specific interest for cultural
heritage. He has been involved in scanning for Sagalassos TR, V&A London, British Museum,
Tongeren, Brussels, Berlin, LA, NY, etc.
2.2.5 Brunswick Town Charitable Trust
Organization profile
Key competences and role in the project
CVs of key personnel
2.3 Consortium as a whole
2.4
Resources to be committed
Proposal Part B: page 29
FP7-ICT-2011-9
18/01/12 v1
3
STREP proposal
SHAPEATLAS
Impact
3.1
Expected impacts listed in the work programme
3.2
Dissemination and/or exploitation of project results, and management of intellectual
property
Proposal Part B: page 30
FP7-ICT-2011-9
18/01/12 v1
4
STREP proposal
SHAPEATLAS
Ethical Issues
YES
Informed Consent
 Does the proposal involve children?
 Does the proposal involve patients or persons not
able to give consent?
 Does the proposal involve adult healthy
volunteers?
 Does the proposal involve Human Genetic
Material?
 Does the proposal involve Human biological
samples?
 Does the proposal involve Human data collection?
Research on Human embryo/foetus
 Does the proposal involve Human Embryos?
 Does the proposal involve Human Foetal Tissue /
Cells?
 Does the proposal involve Human Embryonic
Stem Cells?
Privacy
 Does the proposal involve processing of genetic
information or personal data (eg. health, sexual
lifestyle, ethnicity, political opinion, religious or
philosophical conviction)
 Does the proposal involve tracking the location or
observation of people?
Research on Animals
 Does the proposal involve research on animals?
 Are those animals transgenic small laboratory
animals?
 Are those animals transgenic farm animals?
 Are those animals cloned farm animals?
 Are those animals non-human primates?
Research Involving Developing Countries
 Use of local resources (genetic, animal, plant etc)
 Impact on local community
Dual Use
 Research having direct military application
 Research having the potential for terrorist abuse
ICT Implants
 Does the proposal involve clinical trials of ICT
implants?
I CONFIRM THAT NONE OF THE ABOVE ISSUES
APPLY TO MY PROPOSAL
Proposal Part B: page 31
X
PAGE
FP7-ICT-2011-9
18/01/12 v1
5
STREP proposal
SHAPEATLAS
References
Akbar, S., Kung, J., Wagner, R.: Multi-feature integration with relevance feedback on 3d model
similarity retrieval. J. Mob. Multimed. 3 (2007).
Arandjelovic, R., Zisserman, A.: Ecient image retrieval for 3D structures. In: Proc. BMVC (2010).
Arandjelovic, R., Zisserman, A.: Smooth object retrieval using a bag of boundaries. In: Proc. ICCV
(2011).
Bang, H.Y., Chen, T.: Feature space warping: an approach to relevance feedback. In: Intl. Conf.
Image Proc. Volume 1. (2002).
Bariya, P., Nishino, K.: Scale-hierarchical 3d object recognition in cluttered scenes. In: CVPR
(2010).
Bronstein, A.M., Bronstein, M.M., Guibas, L.J., Ovsjanikov, M.: Shape Google: Geometric words
and expressions for invariant shape retrieval. ACM Trans. Graph (2011).
Chum, O., Perdoch, M., Matas, J.: Geometric min-hashing: Finding a (thick) needle in a haystack.
In: CVPR (2009).
Chum, O., Mikulik, A., Perdoch, M., Matas, J.: Total recall ii: Query expansion revisited. In: CVPR
(2011).
Dutagaci, H., Godil, A., Axenopoulos, A., Daras, P., Furuya, T., R. Ohbuchi, R.: Shrec 2009 shape retrieval contest of partial 3d models (2009).
Efron, M.: Query expansion and dimensionality reduction: Notions of optimality in rocchio
relevance feedback and latent semantic indexing. Inf. Process. Manage (2008).
Elad, M., Tal, A., Ar, S.: Content based retrieval of vrml objects: an iterativeand interactive
approach. In: Proceedings of the sixth Eurographics workshop on Multimedia 2001, New York,
NY, USA, Springer-Verlag (2002).
Fellner D. W., Havemann S., Beckmann P., Pan X.: Practical 3D reconstruction of cultural heritage
artefacts from photographs - potentials and issues. Virtual Archeology Review 2, 4, 95–103 (2010).
Ferrari, V., Zisserman, A.: Learning visual attributes. In: NIPS (2007).
Funkhouser, T.A., Kazhdan, M.M., Shilane, P., Min, P., Kiefer, W., Tal, A., Rusinkiewicz, S.,
Dobkin, D.P.: Modeling by example. ACM Trans. Graph (2004).
Haveman S.: Generative Mesh Modeling. PhD thesis, Institute of Computer Graphics, Faculty of
Computer Science, Braunschweig Technical University, Germany (2005).
Havemann S., Fellner D. W.: Managing procedural knowledge. In Proc. 5th International
Conference on Knowledge Management (I-KNOW’05), Tochtermann K., Maurer H., (Eds.),
Springer, pp. 248–255 (2005).
Havemann S., Fellner D. W.: Patterns of shape design. In Proc. I-KNOW ’09 and I-SEMANTICS
’09, pp. 93–106 (2009).
Proposal Part B: page 32
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
Havemann S., Fellner D. W.: Towards a new shape description paradigm using the Generative
Modeling Language. In Rainbow of Computer Science, Calude C. S., Rozenberg G., Salomaa A.,
(Eds.), vol. 6570 of Lecture Notes in Computer Science, LNCS, Springer, pp. 200–214 (2011).
Hoi, S.C.H., Jin, R.: Active multiple kernel learning for interactive 3d object retrieval systems.
ACM Trans. Interact. Intell. Syst. 1 (2011).
Izadi, S., Kim, D., Hilliges, O., Molyneaux, D., Newcombe, R., Kohli, P., Shotton, J., Hodges, S.,
Freeman, D., Davison, A., Fitzgibbon, A.: Kinectfusion: real-time 3d reconstruction and interaction
using a moving depth camera. In: Proceedings of the 24th annual ACM symposium on User
interface software and technology. UIST '11, New York, NY, USA, ACM (2011).
JSegou, H., Harzallah, H., Schmid, C.: A contextual dissimilarity measure for accurate and efficient
image search. In: CVPR (2007).
JSegou, H., Douze, M., Schmid, C.: On the burstiness of visual elements. In: CVPR (2009).
JSegou, H., Douze, M., Schmid, C.: Improving bag-of-features for large scale image search. IJCV
to appear.
Knopp, J., Prasad, M., Willems, G., Timofte, R., Van Gool, L.: Hough transform and 3d surf for
robust three dimensional classication. In: ECCV (2010).
Knopp, J., Sivic, J., Pajdla, T.: Avoding confusing features in place recognition. In: Proceedings of
the European Conference on Computer Vision (2010).
Leibe, B., Leonardis, A., Schiele, B.: Robust object detection with interleaved categorization and
segmentation. IJCV 77 (2008).
Lowe, D.G.: Distinctive image features from scale-invariant keypoints. IJCV 60 (2004).
Lou, K., Jayanti, S., Iyer, N., Kalyanaraman, Y., Prabhakar, S., Ramani, K.: A reconfigurable 3d
engineering shape search system: Part ii | database indexing, retrieval, and clustering. ASME
Conference Proceedings 2003 (2003).
Mikulik,A., Perdoch, M., Chum, O., Matas, J.: Learning a vocabulary. In: ECCV (2010).
Minka, T.P., Picard, R.W.: Interactive learning with a "society of models". In: Proc. CVPR. CVPR
'96, Washington, DC, USA, IEEE Computer Society (1996).
Ovsjanikov, M., Li, W., Guibas, L., Mitra, N.J.: Exploration of continuous variability in collections
of 3d shapes. ACM Transactions on Graphics (2011).
Pan X., Beckmann P., Havemann S., Tzompanaki K., Doerr M., Fellner D. W.: A distributed object
repository for cultural heritage. In VAST 2010, pp. 105–114 (2010).
Papadakis, I.P.P., T., T., T., T., S., P.: Relevance feedback in content-based 3d object retrieval a
comparative study. In: Computer-Aided Design & Applications (2008).
Philbin, J., Chum, O., Isard, M., Sivic, J., Zisserman, A.: Object retrieval withlarge vocabularies
and fast spatial matching. In: Proc. CVPR (2007).
Proposal Part B: page 33
FP7-ICT-2011-9
18/01/12 v1
STREP proposal
SHAPEATLAS
Philbin, J., Sivic, J., Zisserman, A.: Geometric LDA: A generative model for particular object
discovery. In: Proc. BMVC (2008).
Philbin, J., Sivic, J., Zisserman, A.: Descriptor learning for ecient retrieval. In: ECCV (2010).
Prasad, M., Knopp, J., Van Gool, L.: Class-specific 3D Localization using Constellations of Object
Parts. In: Proc. BMVC (2011).
Qin, D., Gammeter, S., Bossard, L., Quack, T., Gool, L.J.V.: Hello neighbor: Accurate object
retrieval with k-reciprocal nearest neighbors. In: CVPR (2011).
Shilane, P., Min, P., Kazhdan, M., Funkhouser, T.: The princeton shape benchmark. In: Shape
Modeling International (2004).
Sivic, J., Zisserman, A.: Video Google: A text retrieval approach to object matching in videos. In:
ICCV (2003).
Sivic, J., Zisserman, A.: Efficient visual content retrieval and mining in videos. In: Pacific-Rim
Conference on Multimedia, (PCM 2004), Tokyo, Japan (2004).
Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: Exploring photo collections in 3d. In:
SIGGRAPH Conference Proceedings (2006).
Sun, J., Ovsjanikov, M., Guibas, L.: A concise and provably informative multi-scale signature
based on heat diusion. In: SGP (2009).
Toldo, R., Castellani, U., Fusiello, A.: A bag of words approach for 3d object categorization. In:
MIRAGE (2009).
Vergauwen, M., Gool, L.V.: Web-based 3d reconstruction service. Mach. Vision Appl. 17 (2006).
Vranic, D.V., Saupe, D.: 3d shape descriptor based on 3d fourier transform. In: Proceedings of the
EURASIP Conference on Digital Signal Processing for Multimedia Communications and Services
(ECMCS 2001), Budapest, Hungary (2001).
Willems, G., Tuytelaars, T., Van Gool, L.: An ecient dense and scale-invariant spatio-temporal
interest point detector. In: ECCV (2008).
Zhang, W., Surve, A., Fern, X., Dietterich, T.: Learning non-redundant codebooks for classifying
complex objects. In: ICML (2009).
Proposal Part B: page 34
Download