1. TECHNICAL SEMINAR
ON
FAST VISUAL RETRIEVAL
USING ACCELERATED SEQUENCE
MATCHING
PRESENTED BY UNDER THE GUIDANCE
xxxxxxxxx Mr/Mrs.xxxxxxx
(1xx08is041) HOD,DEPT. OF XXX
XXXXXX
2. ABSTRACT
We present an approach to represent, match, and index
various types of visual data.
Primary goal is to enable effective and computationally
efficient searches.
an image/video is represented by an ordered list of feature
descriptors.
Similarities b/w such representations are measured by the
approximate sequence matching technique.
This unifies visual appearance and the ordering
information in a holistic manner.
3. Introduction
With the rapid growth in image/video production and
distribution industry.
Necessary to develop technique which enable users to
easily access and organize large volumes of visual data.
There are 2 major problems regarded to visual retrieval and
its organization i.e.,
(1) lack of design technique for good visual
representation and
(2) lack of quantitative metrics which can efficiently
measure
similarities b/w each pair of visual data.
4. Continued…
The above problems concerned with the visual retrieval
and its organization can be overcomed using following
approaches…
(1) approximate sequence matching technique(also called
Levenshtein distance)
(2)extension to local alignment(smith-waterman
algorithm)
(3) indexing with a Vocabulary tree
(4)Fast matching algorithm
5. 1.Approximate sequence matching
technique
compare two pieces of visual data that represented by
ordered features.
Its is formulated by Levenshtein distance
ci-cost of an operation
The Levenshtein distance is defined as the “minimal cost of
sequence of operations that transforms X into Y”
6. Continued…
• Operations restricted in levenshtein distance are as follows
(1)insertion δ (ε, a)
(2)deletion δ (a, ε)
(3)substitution δ (a, b)
• Levenshtein distance can be computed using dynamic
programming
8. 2.Extension to local alignment
Extended to search for local alignments b/w two feature
sequences.
Derives a score v(xi, yj) between two feature vectors xi and
yj.
v(xi, yj) is +ve if xi and yj are similar.
v(xi, yj) is -ve if xi and yj are not similar.
Value v(xi, yj) is considered as substitution score.
For insertion operation we assign negative scores denoted
as v(xi, ε).
9. Continued…
Similarly for deletion operation we assign scores
denoted as v(ε, yi).
optimal local alignment can be computed using
S(i,j)=max{0,S(i-1,j)+v(x, ε),S(i,j-1)+v(ε,yj),S(i-1,j-1)+v(x
,yj)}
is called Smith-waterman algorithm
compares each descriptor of query to every descriptors
in the database.
10. 3.Indexing with a vocabulary tree
Vocabulary tree-used for indexing all feature vectors
extracted from database.
Indexing each of the visual descriptors can be done
hierarchically using the concept of k-means hierarchical
clustering.
Other advanced vocabulary tree used for indexing
problems is adaptive vocabulary tree
Adaptive vocabulary tree adapts to the addition/deletion of
instances from the database.
Advantage-tree need not to be rebuilt in case of adaptive
vocabulary tree when database undergo slight changes.
It enables the retrieval time to grow sub linearly with
respect to the no of frames in the database.
11. Continued…
Levels of an adaptive vocabulary tree grow sub linearly in
terms of frame numbers in the database
12. 4.Fast matching algorithm
Tries to filter unnecessary alignments which will not lead
to successful matching.
Uses visual method called a dot plot.
Dot plot puts a dot at (i , j) if descriptor i and descriptor j
are similar.
13. Continued…
Diagonal link(i,j) is established if dot j is positioned at the
bottom-right corner of dot i.
Diagonal links represent contiguous matched descriptor
pairs.
Gap link(i,j) is established if dot j is bottom-right
positioned with respect to dot i.
Gap link allows the operation of insertions and deletions.
14. CONCLUSION
Proposed techniques used for effective and computational
efficient visual searches.
Representation based on sequence of features.
We apply the approximate sequence matching method to
measure the similarity based on such a representation.
Presented the framework using which it speedups the
matching process.
Fast retrieval of visuals are ensured.
Good representation of visual data.
Proposed techniques have been demonstrated for use in
several visual retrieval applications.
15. REFERENCES
D. A. Adjeroh, M. -C. Lee, and I. King. A distance measure
for video sequence similarity matching.
M. Bertini, A. D. Bimbo, and W. Nunziati. Video clip
matching using MPEG-7 descriptors and edit distance.
D. Nister and H. Stewenius. Scalable recognition with a
vocabulary tree.
J. Law-To, O. Buisson, V. Gouet-Brunet, and N. Boujemaa.
Robust voting algorithm based on labels of behavior for
video copy detection.