Combining Spanning Trees and Normalized Cuts for Internet Retrieval

1
Combining Spanning Trees and Normalized Cuts for
Internet Retrieval

• Sharat Chandran1 and Abhishek Ranjan2

ViGIL IIT Bombay www.cse.iitb.ac.in/graphics/
2.
1.
2
Need for CBIR

• 1994 Yahoo Text based search
• 2002 Google image search
• search for Apple
• A related problem CBVR

3
Previous efforts

• Image feature vector indexing
• QBIC(1995), Photobook(1996), WBIIS(1997) etc.
• Information loss shape, location etc.
• Image segmentation
• WindSurf(1999), Blobworld(1999), SIMPLIcity(2001)
  etc.
• No iterative refinement

4
An interactive CBIR system

• User enters a query image
• System
• Quickly segments the query image
• Searches images with similar segments
• Returns the approximate results
• User iteratively refines the results
• A progressive refinement strategy
• Our focus Quick segmentation

5
Underlying requirements

• Hierarchical image segmentation
• Control over levels of segmentation
• Fast segmentation for quick response
• Intuitive segmentation

6
Hierarchical segmentation

• Normalized cut (Ncut, Shi et al. 00)
• Global Optimization
• Good criteria
• Promising in videos
• Cost
• Time O(n1.5)
• Space O(n2)
• 106 pixels or 200x100, 50 frames video
• time 1,000,000,000 units,
• space 1,000,000,000,000 units

Image with 2 segments
7
We need

• Speed Quality
• How ?
• Reduce input size fed to algorithm !
• Input size n1/2, Cost O((n1/2)1.5)

8
Two step pipeline
Hierarchical segmentation
Fast grouping
O(n log n)
n9x104
Ncut
O(n1.5)
9
Fast grouping

• Group similar pixels
• Grouping using local variation (PAlgo Pedro et
  al. 1998)
• Uses local properties
• Fast O(n log n)
• Produces too many regions for CBIR

10
Pipeline

• i/p PAlgo intermediate Ncut o/p
• Pipelining not easy
• Unpredictable output size of PAlgo
• Needed output size n1/2
• Keep the quality intact
• Solution
• Cluster merging

11
Merging process

• Sort the similarity
• Merge similar groups iteratively

12
Result
Pipeline O(n log n)
Input
N-cut O(n1.5)
13
Conclusion

• A new pipelining strategy
• Efficiently combined two approaches
• Application in CBIR
• Possibilities
• Video segmentation
• Video retrieval system (CBVR)

14
Thanks!!
Questions ?

aranjan_at_dgp.toronto.edu http//www.dgp.toronto.edu
/aranjan
15
(No Transcript)
16
Why N-cut ?

• Edge weights are proportional to similarity