The document summarizes four presentations from the SIGGRAPH Seminar 2014 session on Shape Collection:
1. "Meta-representations of Shape Families" by Nar Fish et al. which analyzes shape families by computing probability distributions of relations between segmented parts.
2. "Organizing Heterogeneous Scene Collections through Contextual Focal Points" by Kai Xu et al. which extracts focal points from 3D indoor scenes to cluster them.
3. "Geometry and Context for Semantic Correspondences and Functionality Recognition in Man-made 3D Shapes" by Hamid Laga et al. which uses a graph representation and context-aware similarities to find semantic correspondences between parts.
2. Shape Collection
1. Meta-representations of Shape Families
2. Organizing Heterogeneous Scene Collections through
Contextual Focal Points
3. Functional Map Networks for Analyzing and Browsing
Large Shape Collections (unavailable)
4. Geometry and Context for Semantic Correspondences
and Functionality Recognition in Man-made 3D Shapes
5. Learning 3D Attributes of Images through Shape
Collection (unavailable)
Session: Shape Collection
3. Meta-representations of
Shape Families
Nar Fish1, Melinos Averkious2, Oliver van Kaick1,
Olga Sorkine-Hornung3, Daniel Cohen-Or1, Niloy Mitra2
1Tel Aviv University 2University College London 3ETH Zurich
Session: Shape Collection
4. • Analyzing co-segmented 3D shape family by relative
configurations of segments
• Probability distribution of relations
= “identity” of family
Session: Shape Collection
Higher probability
⇒ more valid shape
Meta-representations of Shape Families
N. Fish, M. Averkiou, O. van Kaick, O. Sorkine-Hournung, D. Cohen-Or, N. J. Mitra
5. 1. Abstracting shapes by pre-defined segments
2. Analyzing relations between segments
Session: Shape Collection
computing
convex hull
extracting
OBB
relations: scale, angle, contact
• all pairwise combination
• relative to whole shape
Meta-representations of Shape Families
N. Fish, M. Averkiou, O. van Kaick, O. Sorkine-Hournung, D. Cohen-Or, N. J. Mitra
6. 3. Computing probability distributions
– 1D kernel density estimator (KDE) with common Gaussian
kernel
Session: Shape Collection
(bandwidth setting)
Meta-representations of Shape Families
N. Fish, M. Averkiou, O. van Kaick, O. Sorkine-Hournung, D. Cohen-Or, N. J. Mitra
7. • Exploration of shape families
Session: Shape Collection
Meta-representations of Shape Families
N. Fish, M. Averkiou, O. van Kaick, O. Sorkine-Hournung, D. Cohen-Or, N. J. Mitra
9. Organizing Heterogeneous Scene Collections
through Contextual Focal Points
Kai Xu1,3, Rui Ma2, Hao Zhang2, Chenyang Zhu3,
Ariel Shamir4, Daniel Cohen-Or5, Hui Huang1
1SIAT 2Simon Fraser University 3National University of Defense Technology
4The Interdisciplinary Center, 5Tel Aviv University
Session: Shape Collection
10. • Organizing heterogeneous data (indoor scenes)
– Holistic (singular view) comparison is not meaningful.
(e.g. Paris vs New York)
• Notion of “focal points”
– Representative substructures for attention or focus
– Yielding multiple distance measures depending on FPs
Session: Shape Collection
Organizing Heterogeneous Scene Collections
through Contextual Focal Points
K. Xu, R. Ma, H. Zhang, C. Zhu, A. Shamir, D. Cohen-Or, H. Huang
11. • Input: heterogeneous collection of 3D indoor scenes
– with object labels (bed, table, desk, lamp, chair, etc.)
• Goal: extracting a set of contextual focal points
+ clustering scenes based on the focals
• A contextual focal point is
– Appearing frequently
– Inducing a compact cluster (coherence)
⇒ Focal extraction as optimization
Session: Shape Collection
Organizing Heterogeneous Scene Collections
through Contextual Focal Points
K. Xu, R. Ma, H. Zhang, C. Zhu, A. Shamir, D. Cohen-Or, H. Huang
12. • Iterative co-analysis algorithm
– Frequent pattern analysis
• Exact subgraph isomorphism [Yan & Han 2002]
• Inexact subgraph matching [Riesen et al. 2010]
• Weighted (Cluster-guided) subgraph matching [Tsuda & Kudo 2006]
– Focal-induced scene clustering
Session: Shape Collection
Organizing Heterogeneous Scene Collections
through Contextual Focal Points
K. Xu, R. Ma, H. Zhang, C. Zhu, A. Shamir, D. Cohen-Or, H. Huang
13. • Iterative co-analysis algorithm
– Frequent pattern analysis
– Focal-induced scene clustering
• Clustering on a (BoW feature)
• Subspace segmentation via quadratic programming [Wang et al. 2011]
Session: Shape Collection
Organizing Heterogeneous Scene Collections
through Contextual Focal Points
K. Xu, R. Ma, H. Zhang, C. Zhu, A. Shamir, D. Cohen-Or, H. Huang
14. Session: Shape Collection
Organizing Heterogeneous Scene Collections
through Contextual Focal Points
K. Xu, R. Ma, H. Zhang, C. Zhu, A. Shamir, D. Cohen-Or, H. Huang
15. Geometry and Context for Semantic Correspondences and
Functionality Recognition
in Manmade 3D Shapes
Hamid Laga1, Michela Mortara2, Michela Spagnuolo2
1University of South Australia 2CNR IMATI-Genova
Session: Shape Collection
16. Geometry and Context for Semantic Correspondences and
Functionality Recognition in Manmade 3D Shapes
H. Laga, M. Mortara, M. Spagnuolo
Target: recognizing semantic correspondence
between parts of man-made 3D shapes
– Significant intra-class variations in geometry & topology
⇒ Purely local analysis is useless!
– Goal: unsupervised solution for this problem
Idea: using contextual information (part relations)
– Graph representation & context-aware subgraph similarity
17. • Input: single class 3D shape collection (e.g. vases)
• Output: segmentation w/ semantic correspondence
Algorithm
1. Automatic segmentation of 3D object
• Any algorithm is OK.
2. Constructing graph representation
• Node = part, Edge = structural relationship
• Context of part S = substructure around S
3. Finding correspondence based on similarities on graph
Session: Shape Collection
Geometry and Context for Semantic Correspondences and
Functionality Recognition in Manmade 3D Shapes
H. Laga, M. Mortara, M. Spagnuolo
18. Initial graph construction
• Inter-part symmetries
• Adjacency
• Other contextual relation ships (e.g. enclosure, contact, support)
– Building segmentation hierarchy by clique contraction
Geometry and Context for Semantic Correspondences and
Functionality Recognition in Manmade 3D Shapes
H. Laga, M. Mortara, M. Spagnuolo
Edge
Merge!
Merge!
19. Calculating part-wise correspondence
Session: Shape Collection
Geometry and Context for Semantic Correspondences and
Functionality Recognition in Manmade 3D Shapes
H. Laga, M. Mortara, M. Spagnuolo
Geometric Similarity
Contextual Similarity
p-order similarity function between
Part PA on Graph G1 & Part PB on Graph G2
Compare subgraphs (nodes) by context-aware graph kernel
20. Correspondence results
Session: Shape Collection
Geometry and Context for Semantic Correspondences and
Functionality Recognition in Manmade 3D Shapes
H. Laga, M. Mortara, M. Spagnuolo
21. Functional recognition
– Using graph kernel for supervised learning (SVM)
– Training with labeled 3D objects
– Building binary classifiers
• “Is this part graspable or not?”
Geometry and Context for Semantic Correspondences and
Functionality Recognition in Manmade 3D Shapes
H. Laga, M. Mortara, M. Spagnuolo