Upload
others
View
17
Download
0
Embed Size (px)
Citation preview
Eurographics Workshop on 3D Object RetrievalSGP 2013 Graduate School
Mathematical Tools for 3D Shape Analysis and Description
tools and concepts, part II
Andrea Cerri and Silvia Biasotti
mathematical concepts
– basics on algebraic topology
– simplicial complexes
– homology
– Morse theory
concepts in action
– comparing shapes
– persistent topology
– Reeb graphs (by Silvia)
01/07/2013 3D Shape Analysis and Description
content
2
topological spaces and functions to model
shapes and their properties as pairs 𝑋, 𝑓 ;
suitable for dealing with stability/ robustness;
are there alternatives to isometries to solve
other sets of problems?
01/07/2013 3D Shape Analysis and Description
isometries, but not only…
3
mathematical concepts
– basics on algebraic topology
– simplicial complexes
– homology
– Morse theory
concepts in action
– comparing shapes
– persistent topology
– Reeb graphs (by Silvia)
01/07/2013 3D Shape Analysis and Description
content
4
Associates algebraic invariants with
topological spaces;
Classifies topological spaces up to
homeomorphisms;
Homology is an option.
01/07/2013 3D Shape Analysis and Description
algebraic topology & homology
5
General idea: to distinguish the shape of
topological spaces by their holes;
Several homology theories (singular, cellular,
Čech, etc…);
We will focus on simplicial homology.
01/07/2013 3D Shape Analysis and Description
basics on homology
6
mathematical concepts
– basics on algebraic topology
– simplicial complexes
– homology
– Morse theory
concepts in action
– comparing shapes
– persistent topology
– Reeb graphs (by Silvia)
01/07/2013 3D Shape Analysis and Description
content
7
approach: to decompose a topological
space into simple pieces easier to study;
q-simplices are the building blocks:
– combinatorial aspect: relations among simplices;
– geometric aspect: related to their embedding in
the Euclidean space.
a simplicial complex is the combinatorial
structure, generated by q-simplices;
01/07/2013 3D Shape Analysis and Description
simplicial homology
8
a 𝒒-simplex in ℝ𝑛, with 𝑛 ≥ 𝑞, is the convex
hull of 𝑞 + 1 affinely independent points.
a face of a 𝑞-simplex is the convex hull of a
subset of the set of its points.
01/07/2013 3D Shape Analysis and Description
simplices and faces
0-simplex 1-simplex 2-simplex 3-simplex
9
a (finite) simplicial complex 𝐾 is a (finite)
collection of simplices so that
– if 𝜎 ∈ 𝐾 and 𝜏 is a face of 𝜎, then 𝜏 ∈ 𝐾;
– any two simplices can only meet along a
common face.
01/07/2013 3D Shape Analysis and Description
simplicial complex
10
the dimension of 𝐾 is the highest among the
dimensions of its simplices:
– triangle meshes are 2-complexes;
– tetrahedral meshes are 3-complexes.
01/07/2013 3D Shape Analysis and Description
dimension of a simplicial complex
11
mathematical concepts
– basics on algebraic topology
– simplicial complexes
– homology
– Morse theory
concepts in action
– comparing shapes
– persistent topology
– Reeb graphs (by Silvia)
01/07/2013 3D Shape Analysis and Description
content
12
a 𝒒-chain is a sum of all 𝑞-simplices in 𝐾,
𝜎 = 𝑖 𝜆𝑖 𝜎𝑖, 𝜆𝑖 ∈ ℤ2= {0,1}.
– a curve on a mesh is a 1-chain
– a surface patch is a 2-chain
the 𝒒-chain space 𝐶𝑞 𝐾 is the vector space
generated by all the q -simplices in K ;
01/07/2013 3D Shape Analysis and Description
chains…
13
the boundary operator extracts the
boundary of a chain
the boundary of a boundary is null
01/07/2013 3D Shape Analysis and Description
… and boundaries
0
14
𝐾 is a 𝑘- complex of ℝ𝑛; 𝐶1(𝐾), … , 𝐶𝑘(𝐾) are
the chain spaces. The boundary operator𝜕𝑞: 𝐶𝑞(𝐾) → 𝐶𝑞−1(𝐾)
is defined as:
– for a 𝒒 −simplex 𝜎 = 𝐴0, … , 𝐴𝑞 ,
𝜕𝑞 𝜎 ≝ 𝑖=0𝑞
[𝐴0, … , 𝐴𝑖 , … , 𝐴𝑞] ;
– for chain 𝜎 = 𝑖 𝜆𝑖𝜎𝑖,
𝑖 𝜆𝑖𝜎𝑖, 𝜕𝑞(𝜎) ≝ 𝑖 𝜆𝑖𝜕𝑞(𝜎𝑖).
N.B.: sums are in ℤ2, that is, 1+1=0!
01/07/2013 3D Shape Analysis and Description
boundary operator (more formally)
15
boundary of 𝜎1 = 𝐴0, 𝐴1, 𝐴2 , 𝜕2 𝜎1 and 𝜎 = 𝜎1 + 𝜎2:
𝜕2 𝜎1 = 𝐴1, 𝐴2 + 𝐴0, 𝐴2 + [𝐴0, 𝐴1];
𝜕2 𝜎 = 𝜕2 𝜎1 + 𝜕2(𝜎2) =
= 𝐴1, 𝐴2 + 𝐴0, 𝐴2 + [𝐴0, 𝐴1] + + 𝐴2, 𝐴3 + 𝐴1, 𝐴3 + 𝐴1, 𝐴2 =
= 𝐴0, 𝐴2 + [𝐴0, 𝐴1] + 𝐴2, 𝐴3 + 𝐴1, 𝐴3 .
01/07/2013 3D Shape Analysis and Description
boundary operator (example)
𝜎1
𝜎2
𝐴0
𝐴1
𝐴2
𝐴3
𝜕1 𝜕2(𝜎1) = 𝜕1( 𝐴1, 𝐴2 ) + 𝜕1( 𝐴0, 𝐴2 ) + 𝜕1([𝐴0, 𝐴1])== 𝐴2] + [𝐴1 + 𝐴2] + [𝐴0 + 𝐴1] + [𝐴0 = 0;
16
a 𝑞-chain:
– is a 𝒒-cycle if its boundary is 0;
– is a 𝒒-boundary if is a boundary of a (𝑞 + 1)-chain;
Intuition:
Homology formalizes this idea
01/07/2013 3D Shape Analysis and Description
cycles and boundaries
17
Set 𝒁𝒒 𝑲 = {𝑞-cycles in 𝐶𝑞(𝐾)};
Let 𝑩𝒒 𝑲 = {𝑞-boundaries in 𝐶𝑞(𝐾)};
Since 𝜕 ∘ 𝜕 ≡ 0, then 𝐵𝑞 𝐾 ⊆ 𝑍𝑞 𝐾 ;
The 𝒒-𝒕𝒉 simplicial homology group of 𝐾 is
𝐻𝑞 𝐾 ≝ 𝑍𝑞(𝐾)/𝐵𝑞(𝐾)
01/07/2013 3D Shape Analysis and Description
simplicial homology
18
an element of 𝐻𝑞(𝐾) is an equivalence class
containing homologous 𝑞-cycles;
the rank of 𝐻𝑞(𝐾) is the 𝑞-th Betti number of 𝐾;
the homology 𝐻∗(𝐾) is a topological invariant.
01/07/2013 3D Shape Analysis and Description
simplicial homology (II)
For 3D data:─ 𝛽0 = # components;─ 𝛽1 = # tunnels;─ 𝛽2 = # voids;
19
example: genus of a surface
Let 𝑆 be a connected, orientable surface
without boundary. The genus 𝑔 of 𝑆 is– the maximum number of cuttings along non-
intersecting closed simple curves which can be cut along the surface without disconnecting it…
– … or half of the first Betti number 𝛽1 𝑆 .
The genus is a topological invariant.
3D Shape Analysis and Description01/07/2013 20
mathematical concepts
– basics on algebraic topology
– simplicial complexes
– homology
– Morse theory
concepts in action
– comparing shapes
– persistent topology
– Reeb graphs (by Silvia)
01/07/2013 3D Shape Analysis and Description
content
21
algebraic topology provides invariants up
to homeomorphisms;
what if we want to reason about shapes
under more invariants?
we could use critical points of functions…
01/07/2013 3D Shape Analysis and Description
algebraic topology and differential geometry
22
𝑓: 𝑀 → ℝ smooth function on a smooth
manifold 𝑀 of dimension 𝑛;
– 𝑝 ∈ 𝑀 is critical if 𝑑𝑓𝑝 is the zero map, i.e. 𝑑𝑓
𝑑𝑥1𝑝 = ⋯ =
𝑑𝑓
𝑑𝑥𝑛𝑝 = 0;
– A critical point 𝑝 ∈ 𝑀 is non-degenerate if the Hessian 𝐻𝑓 𝑝 is non-singular, i.e.
01/07/2013 3D Shape Analysis and Description
functions and critical points (I)
det 𝐻𝑓 𝑝 = 𝑑𝑒𝑡𝜕2𝑓
𝜕𝑥𝑖𝜕𝑥𝑗(𝑝) ≠ 0
23
If 𝑝 is a non-degenerate critical point of 𝑓, the index 𝜆𝑝 of 𝑝 is defined as:
𝜆𝑝 = #{𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 𝑒𝑖𝑔𝑒𝑛𝑣𝑎𝑙𝑢𝑒𝑠 𝑜𝑓 𝐻𝑓(𝑝)};
𝑓: 𝑀 → ℝ is a Morse function if all of its critical
points are non-degenerate;
01/07/2013 3D Shape Analysis and Description
functions and critical points (II)
𝜆𝑝 = 0 𝜆𝑝 = 1 𝜆𝑝 = 2
24
𝑝 non-degenerate critical point of index 𝜆𝑝, with
𝑓(𝑝) = 𝑞. Suppose 𝑓−1[𝑞 − 𝜀, 𝑞 + 𝜀] compact, with no critical points besides 𝑝. Then 𝑀𝑞−𝜀 is homotopy
equivalent to 𝑀𝑞+𝜀 with a 𝜆𝑝-dim. disk attached.
For 𝑐 ∈ ℝ , set 𝑀𝑐 = 𝑝 ∈ 𝑀: 𝑓 𝑝 ≤ 𝑐 = 𝑓−1( −∞, 𝑐 ).
How/when does the topology of 𝑀𝑐 changes?
Let 𝑎 < 𝑏. Suppose 𝑓−1 𝑎, 𝑏 = 𝑝 ∈ 𝑀: 𝑎 ≤ 𝑓 𝑝 ≤ 𝑏is compact and has no critical points. Then 𝑀𝑎 and
𝑀𝑏 are diffeomorphic.
01/07/2013 3D Shape Analysis and Description
Morse functions & critical points: Two results
25 01/07/2013 3D Shape Analysis and Description
Morse functions & critical points: Example
𝛽0 𝛽1 𝛽226
Morse theory can be used for..
to combine the topology of 𝑀 with the
quantitative measurement provided by 𝑓
– 𝑓 is the lens to look at the properties of (𝑀, 𝑓)
– different choices of 𝑓 provide different invariants
to construct a general framework for shape
characterization which if parameterized wrt
the pair (𝑋, 𝑓)
3D Shape Analysis and Description01/07/2013 27
S. Willard, General Topology, Addison-Wesley Publishing Co, 1970
R. Engelking and K. Sielucki, Topology: A geometric approach, Sigma series in pure mathematics, Heldermann, Berlin, 1992
A. Fomenko, Visual Geometry and Topology, Springer-Verlag, 1995M.
W. Hirsch, Differential Topology, Springer, 1997
H. B. Griffiths, Surfaces, Cambridge University Press, 1976
V. Guillemin and A. Pollack, Differential Topology, Englewood Cliffs, NJ:Prentice Hall, 1974
W. Massey, Algebraic topology: An Introduction, Brace&World Inc., 1967
A. Hatcher, Algebraic Topology, Cambridge University Press, 2001
J. Milnor, Morse theory, Princeton University Press, New Jersey, 1963
C. Kosniowski, A First Course in Algebraic Topology, Cambridge University Press, 1966
01/07/2013 3D Shape Analysis and Description
references
28
mathematical concepts
– basics on algebraic topology
– simplicial complexes
– homology
– Morse theory
concepts in action
– comparing shapes
– persistent topology
– Reeb graphs (by Silvia)
01/07/2013 3D Shape Analysis and Description
content
29
Comparing objects’ shapes can done with respect
to some “relevant properties”;
To model a shape we consider pairs (𝑋, 𝑓) s.t.
– X represents the object;
– 𝑓: 𝑋 → ℝ is a function and describing the relevant
properties of the object.
01/07/2013 3D Shape Analysis and Description
comparing shapes (I)
(𝑋, 𝑓) (𝑋, ℎ)(𝑋, 𝑔) (𝑋, 𝑙)
30
How can we compare two pairs (𝑋, 𝑓), (𝑌, 𝑔) ?
Idea: Use a metric able to quantify and measure the deformation of 𝑋 into 𝑌, in terms of 𝑓 and 𝑔.
comparing shapes (II)
(𝑋, 𝑓) (𝑌, 𝑔)
01/07/2013 3D Shape Analysis and Description 31
formally: Natural pseudo-distance 𝒅
𝑑( 𝑋, 𝑓 , 𝑌, 𝑔) ≝ infℎ
max𝑥∈𝑋
𝑓 𝑥 − 𝑔 ℎ 𝑥 ∞ ,
ℎ varying among all homoeomorphisms from X to Y.
natural pseudo-distance (I)
(𝑋, 𝑓) (𝑌, 𝑔)
01/07/2013 3D Shape Analysis and Description 32
Problem: the natural pseudo-distance 𝑑 is difficult to
compute; we need tools to get information about it;
Persistent topology allows us get lower bounds for 𝒅by means of suitable shape descriptors;
Instead of comparing pairs (space,functions), we can compare the associated descriptors.
natural pseudo-distance (II)
01/07/2013 3D Shape Analysis and Description 33
mathematical concepts
– basics on algebraic topology
– simplicial complexes
– homology
– Morse theory
concepts in action
– comparing shapes
– persistent topology
– Reeb graphs (by Silvia)
01/07/2013 3D Shape Analysis and Description
content
34
a filtration is a nested sequence 𝑋1 ⊆ 𝑋2 ⊆ ⋯𝑋𝑛 = 𝑋,
e.g. the sub-level sets of a function 𝑓: 𝑋 → ℝ.
topological exploration along a filtration of X, looking for important topological events.
persistence: Filtrations
01/07/2013 3D Shape Analysis and Description 35
Measure the lifespan of homology classes along the
filtration:
Encode the birth level 𝑖 and the death level 𝑗 of a homology class by a point (𝑖, 𝑗).
persistence diagrams (I)
01/07/2013 3D Shape Analysis and Description
𝑓
𝑖
𝑗
36
01/07/2013 3D Shape Analysis and Description
persistence diagrams (II)
𝑓𝑋
𝛽0 𝛽1 𝛽2
37
Changing functions produces different information.
01/07/2013 3D Shape Analysis and Description
persistence diagrams: Modularity
38
Invariance properties w.r.t. transformation groups
are inherited from the function.
01/07/2013 3D Shape Analysis and Description
persistence diagrams: Invariance properties
39
Stability with respect to the bottleneck distance*:
𝑑𝐵 Dgm 𝑓 , Dgm 𝑔 ≤ infℎ
max𝑥∈𝑋
𝑓 𝑥 − 𝑔 ℎ 𝑥∞
This implies resistance to noise;
It is a lower bound for the natural pseudo-distance.
________________* Cohen-Steiner et al. (2005), Chazal et al. (2009), d’Amico et al. (2010)...
01/07/2013 3D Shape Analysis and Description
persistence diagrams: stability
(𝑋, 𝑓) (𝑌, 𝑔)
40
system for content-based retrieval of figurative;
Dataset containing more than 10000 trademarks;
Goal: For an image, find in the dataset the most similar ones.
01/07/2013 3D Shape Analysis and Description
Retrieval of trademark images [C., Ferri, Giorgi 2006]
41
Approach:
– for each image, battery of 25 descriptors;
– each distance between descriptors of the same type induces a metric on the database;
– combine all 25 metrics (max, sum…) to obtain a
final similarity score.
01/07/2013 3D Shape Analysis and Description
Retrieval of trademark images [C., Ferri, Giorgi 2006]
42
Point cloud data analysis [Chazal et al. ‘09]
3D segmentation [Skraba et al. ‘10]
01/07/2013 3D Shape Analysis and Description
Other possible applications…
43
data simplification [Bauer, Lange, Wardetzky ‘12]
3D (textured) shape analysis and comparison [Biasotti et al. ‘13]
01/07/2013 3D Shape Analysis and Description
Other possible applications…
44
Size functions [Frosini ‘91];
Persistent homology groups [Edelsbrunner, Letscher, Zomorodian
‘02];
Vines and vineyards [Cohen-Steiner, Edelsbrunner, Morozov ‘06];
Interval persistence [Dey, Wegner ‘07];
Zig-Zag persisetnce [Carlsson, de Silva, Morozov ‘09];
Persistent cohomology [de Silva, Morozov, Vejdemo-Johansson];
…
Multidimensional setting (𝑓: 𝑋 → ℝ𝑘):
multidimensional persistent homology group (Carlsson, Zomorodian ‘07);
Multidimensional size functions (Biasotti et al. ‘08);
Persistence spaces (C., Landi ‘13);
…
01/07/2013 3D Shape Analysis and Description
an historical perspective
45
Reeb graph
Reeb graphs store the evolution of the level
sets of the mapping function f
M
f
3D Shape Analysis and Description01/07/2013 46
let M be a compact n-manifold, f: M→R a simple Morse function, and given “~” the
equivalence relation
(P, f(P)) ~ (Q, f(Q)) f(P) = f(Q) and P and Q are in the same connected component of 𝑓−1(𝑓(𝑃))
the quotient space on Mxℝ is a finite, connected simplicial complex K of dimension 1,
such that
– the counter-image of each vertex of K is a singular
connected component of the level sets of f
– the counter-image of the interior of each simplex of
dim 1 is homeomorphic to the topological product of one connected component of the level sets by ℝ
01/07/2013 3D Shape Analysis and Description 47
Reeb graph definition
01/07/2013 3D Shape Analysis and Description
overview of RGs when the function f varies
height bounding sphere
center
integral
geodesic
curvature
extrema
barycenter
f values
48
quiz
draw the Reeb graph with respect to the
height function f of the following shapes
f
3D Shape Analysis and Description01/07/2013 49
quiz - solutions
f
3D Shape Analysis and Description01/07/2013 50
01/07/2013 3D Shape Analysis and Description
RG properties
it provides a 1D structure of the shape
it describes the shape of an object under
the lens of the function f
nodes and arcs depend on the number of
critical points of f
it may be computed in operations )log( nnO
51 01/07/2013 3D Shape Analysis and Description
Reeb graph based representations
Reeb graph variations
– contour trees (simply-connected domains)
– component trees (gray-level images)
– centerline skeletons (geodesic distance from a point)
for shape matching
– Multiresolution Reeb graph (MRG), Hilaga et al. 2001
– augmented Multiresolution Reeb graph (aMRG), Tung&Schmitt 2005
– Extended Reeb graph (ERG), Biasotti et al. 2000
52
extension to volume data
f
distance
from the
barycentre
01/07/2013 3D Shape Analysis and Description 53
geometric embedding
nodes of the ERG correspond to regions (eithersurfacic or volumic)
arcs code the adjacency among the parts
each arc can be oriented using the growing direction of the mapping function: the RG is a direct acyclic graph
with each ERG node, store spatial attributes measuring properties of the part associated to it– node embedding (Cartesian coordinates, x, y, z)
– average value of f in the part
– volume
– area
– some radii: min, max, average
01/07/2013 3D Shape Analysis and Description
rmin
ravrmax
54
geometric embedding
each arc can be oriented using the increasing direction of f: the RG is a direct acyclic graph
nodes and arcs correspond to surface parts
node and arc attributes are stored in terms of the spherical harmonic indices of the parts
01/07/2013 3D Shape Analysis and Description 55 01/07/2013 3D Shape Analysis and Description
part correspondence
models with similar appearance
56
01/07/2013 3D Shape Analysis and Description
part correspondence
objects with dissimilar global appearance
57
shape retrieval
performance on the SHREC'07 watertight
database
01/07/2013 3D Shape Analysis and Description
ERG
ERG
ERG
58
shape approximation and compression
given the Reeb graph RG (M,f) of f Morse
and simple, a rough shape chartification is
obtained
1-strip
contains one max or min
one boundary component
3-strips
contains one saddle
three boundary components
f may assume different values
on each of these three
boundary components
RG (M,f)
01/07/2013 3D Shape Analysis and Description 59
v=3.8K
138K
v=35K
t =69K
2,6MB
01/07/2013 3D Shape Analysis and Description 60
references
P. Frosini, A distance for similarity classes of submanifolds of a Euclidean space, Bull. Australian Mathematical Society 42 , 407–416, 1990
P. Frosini, Measuring shapes by size functions, SPIE, Intelligent Robots and Computer Vision X, Vol. 1607, pp. 122-133, 1991
P. Frosini and C. Landi, Size functions and formal series, Appl. Algebra Engrg. Comm. Comput., Vol. 12, pp. 327-349, 2001
H. Edelsbrunner, J. Harer. Computational Topology - an Introduction. AMS , I-XII, 1-24 1, 2010
H. Edelsbrunner , J. Harer, Persistent homology: a survey. Surveys on discrete and computational geometry, vol. 453 of Contemp. Math AMS, 2008
R. Ghrist. Barcodes: the persistent topology of data, Bull. Amer. Math. Soc., 45(1):61-75, 2008
S Biasotti, A Cerri, P Frosini, D Giorgi, C Landi. Multidimensional size functions for shape comparison, J. of Math. Imaging and Vision 32 (2), 161-179
S. Biasotti, A. Cerri, P. Frosini, D. Giorgi, A new algorithm for computingthe 2-dimensional matching distance between size functions, Patt. Rec. Letters 32 (14), 1735-1746
3D Shape Analysis and Description01/07/2013 61
references
G. Reeb, Sur les point sunguilèrs d’une form del Pfaff complètement intégrable oud’une fonction munèrique, Comptes Rendu de l’Academie des Sciences, 222:847-849, 1946
Y. Shinagawa, T. Kunii, Y. Kergosien, Surface coding based on Morse Theory, IEEE Computer Graphics and Applications, 11(5):66-78, 1991
M. Hilaga, Y. Shinagawa, T. Komura, T. Kunii, Topology matching for fullyautomatic similarity estimation of 3D shapes, Proc. SIGGRAPH 2001, pp. 203-212, 2001
S. Biasotti, L. De Floriani, B. Falcidieno, P. Frosini, D. Giorgi, C. Landi, L. Papaleo,M. Spagnuolo, Describing shapes by geometrical-topological properties of real functions, ACM Computing Surveys, 40(4):1-87, 2008
S. Biasotti, D. Giorgi, M. Spagnuolo, B. Falcidieno, Reeb graphs for shape analysis and applications, Theoretical Computer Science, Elsevier, 392(1-3):5-22, 2008
S. Biasotti, G. Patanè, M. Spagnuolo, B. Falcidieno, G. Barequet, Shape approximation by differential properties of scalar functions, Computers & Graphics, 34(3):252-262, 2010
T. Dey, Y. Wang. Reeb graphs: approximation and persistence, Proc. Symposium comput. Geom., 226-235, 2011
V. Barra, S. Biasotti, 3D shape retrieval using Kernels on Extended Reeb Graphs, Pattern Recognition, 46(11):2985-2999, 2013
3D Shape Analysis and Description01/07/2013 62
Eurographics Workshop on 3D Object RetrievalSGP 2013 Graduate School
Mathematical Tools for 3D Shape Analysis and Description
wrap up
Silvia Biasotti
to sum up: theory says …
we have seen:– how to represent a shape (topological spaces,
manifolds, metric spaces, simplicial complexes…)
– how to analyze, describe, compare shapes(geodesics, curvature, diffusion geometry, Morse theory, algebraic machinery…)
– examples of different applications
we have investigated:– space, invariance, property…
– theoretical assumptions vs real world assumptions
but if we go back to «real» shapes and problems, are we sure that everything works ok?– services which could use the theories discussed
– discretization issues and applicability
01/07/2013 3D Shape Analysis and Description 64
complex shapes, complex problems
online repositories of 3D models
– Google 3D Warehousehttp://sketchup.google.com/3dwarehouse/
01/07/2013 3D Shape Analysis and Description 65
complex shapes, complex problems
online repositories of 3D models
– Google 3D Warehousehttp://sketchup.google.com/3dwarehouse/
– 3Dvia http:// www.3dvia.com/search/
01/07/2013 3D Shape Analysis and Description 66
complex shapes, complex problems
online repositories of 3D models
– Google 3D Warehousehttp://sketchup.google.com/3dwarehouse/
– 3Dvia http:// www.3dvia.com/search/
– Turbosquid http://www.turbosquid.com/
• “…search our stock catalog to get the 3D
model you want, or use our Custom 3D modeling service for made-to-order 3D
models. Join the world's top artists who use TurboSquid 3D models in advertising, architecture, broadcast, games, training, film,
the web, and just for fun”
01/07/2013 3D Shape Analysis and Description 67
complex shapes, complex problems
3D Object Retrieval & Content-based
search
– Princeton Shape Benchmark
– Digital Shape Workbench v5.0
http://visionair.ge.imati.cnr.it/
01/07/2013 3D Shape Analysis and Description 68
to sum up: theory says …
3D shapes maybe very complex and serviceswe can imagine for sharing, searching, and
even design new shapes are many, so…
… how to get out of this maze?
basics
– the choice of one or another shape description
should be guided by
• type of shapes and their variability/complexity
• invariants or properties to be preserved / captured
– topological spaces and functions are promising to
model and reason about similarity in a mathematically well-formulated manner
01/07/2013 693D Shape Analysis and Description
pay attention to…
… the right space– rigid transformations (rotations, translations)
• Euclidean distances
– isometries
• Riemannian metric
• curvature (but unstable to local noise/perturbations)
• geodesics, diffusion geometry, Laplacian operators, etc
– local invariance to shape parameterizations
• conformal geometry
– similarities (i.e. scale operations)
• normalized Euclidean distances
– affinity (and homeomorphisms)
• persistent topology
• Morse theory
• size theory
7001/07/2013 3D Shape Analysis and Description
pay attention to…
… a suitable shape description– coarse coding (but fast)
• histograms
• matrices
– articulated shapes
•medial axes
• Reeb graphs
– overall global appearance
• silhouettes
– if shape loops are relevant
• persistent topology
• graph-based descriptions
7101/07/2013 3D Shape Analysis and Description
to sum up: benchmarking
importance of benchmarking and evaluation!– not only publishing papers but also to demonstrate
real innovation and applicability potential
Shape segmentation benchmark
SHape Retrieval Contexthttp://www.aimatshape.net/event/SHREC– an annual event to to evaluate the effectiveness of
3D shape analysis algorithms
– a multi-track event spanning
• different models: from watertight objects to triangle soups, from abstract shapes to medicaldata
• different tasks: from 3D retrieval to correspondence finding and segmentation
7201/07/2013 3D Shape Analysis and Description
example: SHREC'12 - SAS
to evaluate the stability of the algorithms
under variations of abstract shapes
characterized by
– smooth/sharp features,
– partial/global symmetries
– different genus
to create a new benchmark where to
measure the capability of the retrieval
methods to be invariant to
– noise addition and resampling
– deformations
37301/07/2013 3D Shape Analysis and Description
dataset
4
1. addition of Gaussian noise
2. different regular sampling
3. uneven sampling
4. & .5 two stretching
6. & .7 two non-uniform dilations
8. & .9 two non-uniform erosions
9 transformations, 3 degrees each, for a total of 27 perturbations
7401/07/2013 3D Shape Analysis and Description
dataset
47501/07/2013 3D Shape Analysis and Description
504 watertight models without self-intersections
57601/07/2013 3D Shape Analysis and Description
recall-precision curves (whole dataset)
01/07/2013 3D Shape Analysis and Description 77
recall-precision curves (deformations)
01/07/2013 3D Shape Analysis and Description 78
conclusions
the right tool for the right problem
– mathematically sound methods
– benchmarking & evaluation
geometry, structure, similarity, context
– is it possible to understand something about
functionality or affordance?
– machine learning vs geometric-reasoning
– 3D query modalities
– what if shape is influenced/modified by the context?
01/07/2013 3D Shape Analysis and Description 79
acknowledgements
IQmulus: A High-volume Fusion and Analysis Platform for Geospatial Point Clouds, Coverages and Volumetric Data Sets, European project “FP7 IP”, 2012-2106
VISIONAIR: Vision Advanced Infrastructure for Research, European project “FP7 INFRASTRUCTURES”, 2011-2015
MULTISCALEHUMAN: Multi-scale BiologicalModalities for Physiological Human Articulation, European project “FP7 PEOPLE” Initial Training Network, 2011-2014
Patrizio Frosini , Massimo Ferri and the Vision Mathematics group at the Dept. of Mathematics, University of Bologna
Claudia Landi, Dept. of Science and Methods of Engineering, University of Modena and Reggio Emilia
8001/07/2013 3D Shape Analysis and Description
Eurographics Workshop on 3D Object RetrievalSGP 2013 Graduate School
at the end…
thank you for your attention!
these course notes are available at: http://www.ge.imati.cnr.it/training