Spatial Bayesian Nonparametrics for Natural Image Segmentation
Spatial color image retrieval without segmentation using ... · Spatial color image retrieval...
-
Upload
truongtruc -
Category
Documents
-
view
229 -
download
2
Transcript of Spatial color image retrieval without segmentation using ... · Spatial color image retrieval...
Spatial color image retrieval without segmentation using thumbnails and the Earth Mover’s Distance
Thomas Hurtut, Haroldo Dalazoana, Yann Gousseau, Francis Schmitt
CGIV 2006 – Leeds – 20 June 2006
1
1. Introduction2. Features3. Distance4. Matching criterion5. Experimental evaluation6. Conclusion
Presentation Plan
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
2
Color image retrieval:
Introduction
Unavoidable segmentation errors false retrievals
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
without spatial information
with spatial information
3
Color-spatial image retrieval, many approaches:- augmented features: e.g. chromatic histograms [Lambert et al. CGIV 2004]- points of interest : e.g. [Heidemann 2004]- region based image retrieval (RBIR): e.g. BlobWorld [Malik et al. 2002],
[Rugna et al. CGIV 2002]
( )
Coarsely sampled thumbnails, size n pixels
Features
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
Signature: n features f i
4
Earth Mover Distance (EMD) is used [Rubner et al. 2000].
Distance
Solved with Kuhn-Munkres (1955) algorithm for assignement problems in O(n3)
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
General case (transportation problem): Assignement problem case
Best solved with simplex algorithm (can be exponential with n)
Our features/segmentation step takes full advantage of EMD
5
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
Unsupervised matching criterion
An a contrario model gives an unsupervised answer: we do perceiveevents that happen rarely in a noisy situation.
Given a query, which images from the database are relevant ?
A target image will be considered as similar to a query if it is closer to the query than an image of « noise » could be.
A few images...
A littlle more...
Much more !
6
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
Unsupervised matching criterion
The noise model we consider is a dead leaves model [Matheron 1968], consisting in the superimposition of random objects, together with a power law distribution for the size of objects.
The distribution of the color of objects is learned from the color distribution of pixels from thumbnails of the database.
7
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
Experimental evaluationTests on a illuminated manuscripts database from IRHT gathering 1500 images
Blobworld + EMD Our method
8
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
Experimental evaluationTests on a illuminated manuscripts database from IRHT gathering 1500 images
9
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
Experimental evaluationTests on a illuminated manuscripts database from IRHT gathering 1500 images
10
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
Experimental evaluationTests on a database gathering 25000 images from multiple Canadian museums
11
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
Experimental evaluation 12
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
Experimental evaluationTests on a database gathering 25000 images from multiple Canadian museums
13
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
Conclusion and future work
ConclusionWe propose a spatial color image retrieval method without initial
segmentation, using thumbnails and EMD:- robust description- efficient use of the EMD distance
We also propose an unsupervised matching criterion relying on an a contrario method.
Future workMore elaborated features: local quantization, texture characteristicsTests on bigger databases (> 100 000 images)Recent EMD approximation, to speed up querying
14
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
Questions ?
Acknowledgments
We thank Gilles Kagan from IRHT for providing us the IRHT database and for his helpful comments. We also thank Farida Cheriet from LIV4D Polytechnic School of Montreal for her suggestions. This work was supported by a CNRS ACI supervised by Anne-Marie Edde and Dominique Poirel from IRHT, and by a grant from the LIV4D laboratory.
15
More results are visible at: http://www.tsi.enst.fr/recherche/cbir
Distance
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
balances color and spatial distribution
tune the color dynamic,
tune the spatial dynamic
16
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
Experimental evaluationTests on a illuminated manuscripts database gathering 1500 images
balances color and spatial distribution
17
(color only)
(color and spatial)
Introduction Features Distance Unsupervised matching criterion Experimental evaluation Conclusion
Unsupervised matching criterion
Given an image database
We make a statistical test of two hypothesis against each other:
And we will say that:
Where is such that:
18
The expected number of false matches is one