3D Object Retrieval: history and future challenges Benjamin Bustos CIWS / PRISMA Research Group Department of Computer Science University of Chile
[email protected]
Motivation
Lots of practical applications for this problem!
2
Motivation
Main problems for solving 3D object retrieval
Representation of the 3D model
How to measure similarity?
Objects at different scales / “resolution” / pose Triangle meshes / Point clouds Quality (holes, badly oriented triangles, etc.) Geometric or semantic?
What about partial similarity?
Whole-to-part / Part-to-whole
3
Outline
Global descriptors Local descriptors Evaluation: SHREC Tracks Future challenges
4
GLOBAL DESCRIPTORS
5
3D model similarity search
Geometric or semantic similarity?
6
3D model similarity search
General approach: feature vectors
7
3D model similarity search
Requirements for 3D descriptors
Invariance with respect to rotations, translations, scaling and (in some cases) reflections
8
3D model similarity search
Requirements for 3D descriptors
Robustness with respect to level-of-detail (different tessellations)
9
3D model similarity search
Requirements for 3D descriptors
Robustness with respect to outliers Efficient feature extraction Multiresolution representation Discrimination of different geometric forms
10
Normalization of 3D objects
Goal
Result
Find a canonical coordinate system Invariance against translation, rotation, scaling and reflections
Tools*
Center of mass (translation) Continuous PCA (rotation) Test based on moments (reflection) Scaling factor
* D. Vranic, D. Saupe, and J. Richter. Tools for 3D-object Retrieval: Karhunen-Loeve Transform and Spherical Harmonics. In Proc. IEEE Workshop Multimedia Signal Processing, pages 293—298, 2001
11
Normalization of 3D objects
Global transformation of the object with PCA
12
Normalization of 3D objects
Examples of pose normalization with PCA
13
3D descriptors
Extracting 3D feature vectors*
* B. Bustos, D. Keim, D. Saupe, T. Schreck, and D. Vranic. Feature-based similarity search in 3D object databases. ACM Computing Surveys, 37(4):345-387, 2005
14
Shape distribution with D2*
Category: Statistical
Select two random points on the surface Compute Euclidean distance between the points Repeat the process and construct a histogram of distances
* R. Osada, T. Funkhouser, B. Chazelle, and D. Dobkin. Shape distributions. ACM Transactions on Graphics, 21(4):807—832, 2002
15
Extended gaussian image*
Category: surface geometry
* B. Horn. Extended Gaussian Image. Proceedings of the IEEE, 72(12):1671—1686, 1984 16
Ray-based*
Category: extension-based (dim=162)
* D. Vranic and D. Saupe. 3D Model Retrieval. In Proc. Spring Conference on Computer Graphics and its Applications (SCCG’00), pages 89—93, 2000
17
Volume-based FV*
Category: volume-based
* M. Heczko, D. Keim, D. Saupe, and D. Vranic. Methods for similarity search of 3D objects. Datenbank-Spektrum, 2(2):54—63, dpunkt.verlag, 2002
18
Depth-buffer FV*
Category: image-based
Apply 2D discrete Fourier transform to the six images
* M. Heczko, D. Keim, D. Saupe, and D. Vranic. Methods for similarity search of 3D objects. Datenbank-Spektrum, 2(2):54—63, dpunkt.verlag, 2002
19
Experimental evaluation
Experimental framework*
1,838 3D objects collected from the web
472 classified objects in 55 classes (planes, cars, animals, swords, plants, human figures, etc.) Konstanz Dataset
16 implemented 3D descriptors Effectiveness evaluation measures
Precision vs. recall diagrams R-precision
* B. Bustos, D. Keim, D. Saupe, T. Schreck, and D. Vranic. An experimental effectiveness comparison of methods for 3D similarity search. International Journal on Digital Libraries, Special issue on Multimedia Contents and Management in Digital Libraries, 6(1):39-54, 2006
20
Experimental evaluation
Average precision vs. recall
21
Experimental evaluation
Dependency on dimensionality
22
Experimental evaluation
Dependency on query object
23
Experimental evaluation
Dependency on query object
24
Disadvantages of global descriptors
Many other global descriptors:
But:
Image-based: best in practice Combinations of descriptors (fixed, dynamic) Global descriptors suffer from pose sensitivity They are not well-suited for partial similarity
New approach: Local descriptors!
Avoid the canonical pose problem Support non-rigid and partial matching 25
LOCAL DESCRIPTORS
26
Local descriptors for 3DOR
Main approach
Each shape is represented as a set of descriptors The matching is performed by finding correspondences
Example: Shape Google *
Heat kernel signatures Bag-of-features approach
* Ovsjanikov, M., Bronstein, A.M., Guibas, L.J., Bronstein, M.M.: Shape Google: a computer vision approach to invariant shape retrieval. In: Proc. Workshop on Non-Rigid Shape Anal. and Deform. Image Alignment, 2009
27
Key-points
Interest points on the 3D surface
Harris 3D method *
* I. Sipiran and B. Bustos. Harris 3D: A robust extension of the Harris operator for interest point detection on 3D meshes. The Visual Computer, 27(11):963-976, 2011
28
Key-points
Harris 3D is a robust interest points detector
29
Key-components
Regions of interest on the 3D shape *
Based on key-points
* I. Sipiran and B. Bustos. Key-components: Detection of salient regions on 3D meshes. The Visual Computer 29(12):1319-1332, 2013
30
Key-components
Main idea: clustering of key-points in geodesic space
31
Key-components
Key-components are robust
32
Finding correspondences *
* I. Sipiran and B. Bustos. A fully hierarchical approach for finding correspondences in non-rigid shapes. In Proc. 14th IEEE International Conference on Computer Vision (ICCV'13), pages 817-824. IEEE, 2013
33
Finding correspondences
Pre-processing
Key-point detection: Harris 3D algorithm Compute geodesic distances between every pair of key-points Discard key-points with low density For each remaining key-point, compute descriptors
Internal nodes: Heat Kernel Signature [Sun et al. 2009] Leaves: Wave Kernel Signature [Aubry et al. 2011]
34
Finding correspondences
Decomposition tree
35
Finding correspondences
Decomposition tree: node creation
Compute smallest sphere that encloses keypoints within the cluster Find the medoid of the cluster Extract mesh patch (key-component) Compute descriptor for the region
Average key-points descriptor
Continue decomposition if possible
Area patch is large It contains enough key-points 36
Finding correspondences
Hierarchical matching
Given two shapes and their corresponding decomposition trees, matching is performed levelwise
37
Finding correspondences
SHREC'2010 correspondence dataset
It evaluates the robustness of correspondences to transformations: holes, noise, topology changes, etc. Data:
Three shapes (null shapes) Set of transformed versions of the null shapes
45 transformed shapes for each null shape
Measure: geodesic error
38
Experimental evaluation
Some results
39
Experimental evaluation
Strong perturbation of the data
40
EVALUATION: SHREC TRACKS
41
SHREC tracks
Objective: “evaluate the effectiveness of 3Dshape retrieval algorithms” * Started in 2006 It provide many resources for evaluating 3D retrieval methods:
Reference collections Descriptions of the algorithms But no open source code (yet)
* http://www.projects.science.uu.nl/shrec/
42
SHREC tracks
Current typical tracks:
Non-rigid shape matching 3D sketch-based retrieval Retrieval based on range scans or depth scanners Multimodal 3D object retrieval
43
SHREC tracks
Typical evaluation measures:
Mean Average Precision Nearest Neighbor First Tier Second Tier Discounted Cumulative Gain
44
SHREC Track 2013
Large-Scale Partial Shape Retrieval Track Using Simulated Range Images
Task: partial retrieval (part-to-whole) Queries were simulated as if they were taken using an RGB-D camera Dataset:
360 models (20 classes, 18 objects per class) Queries: 20 partial views per model, 7.200 queries
45
SHREC Track 2013
Object classes:
46
SHREC Track 2013
Query simulation:
47
SHREC Track 2013
Ten teams registered for the track… … but only two teams submitted results!
Some teams had problems regarding the size of the dataset Other teams had problems processing the simulated range scans
48
SHREC Track 2013
Main results
49
SHREC Track 2013
Conclusions
Overall performance was moderate: problem far from being solved Efficiency and robustness issues
How to deal with noisy or “imperfect” data? How to do it fast?
50
SHREC Track 2015
Scalability of 3D Shape Retrieval Task: non-rigid shape retrieval Dataset:
229 non-rigid shapes, 9 classes [Summer]
Plus objects from other well known datasets Plus parts of objects using the simulated range scans (post-processed)
51
SHREC Track 2015
Our dataset: 96,487 models
Previous largest dataset: 8,987 models
52
SHREC Track 2015
Four teams participated in the track We wanted to evaluate efficiency
First time that information about the hardware was requested to teams
For evaluation of effectiveness, we used the standard measures (NN, FT, ST, MAP)
53
SHREC Track 2015
Main results: effectiveness
54
SHREC Track 2015
Conclusions:
Near-perfect effectiveness with one of the methods Efficiency vs. effectiveness trade-off Effective methods are not scalable (yet)
55
FUTURE CHALLENGES
56
Future challenges
Partial matching “Functional” matching Scalability issues
Preprocessing steps Computation of the 3D descriptors Similarity search
57
Future challenges
Similarity search using point cloud data directly obtained from sensors Multimodal search
58
Thank you for your attention!
59