Image Database Design Based on 9DSPA Representations for Spatial Relations

1
Image Database Design Based on 9D-SPA
Representations for Spatial Relations
Paper review
Po-Whei Huang, Chu-Hui Lee. IEEE Transactions on
Knowledge and Data Engineering, vol. 16, no. 12,
pp. 1486-1496, December, 2004.

• James M. Kang
• Liang Guo
• 8715 Spatial Databases
• http//www-users.cs.umn.edu/jkang/8715/

2
Outline

• Motivation Problem Statement
• Contributions
• Key Concepts
• Validation Methodologies
• Assumptions
• Rewrite today

3
Motivation

• Encoding the spatial relations in an image to
  facilitate content based image retrieval in a
  pictorial database.
• Pictorial database Spatial relations
• Content based image retrieval
• Finding all the pictures containing a house to
  the west of a car

4
Motivation

• Application domain
• GIS
• CAD
• Office Automation
• Medical image archiving
• Trademark picture registration
• Applied to any image database domain
• Incl. Photographs

5
Problem Statement

• Given
• A set of pictures P, each p containing n objects
  (O1, O2, .., On) where each o corresponds to an
  icon i
• Find
• A set of similar images based on a query image
• Objective
• Minimize computational time to query an image
  index
• Constraints
• Objects in an image are pre-determined
• Objects are assigned to a pre-determined Icon
• The number of Icons in the database are fixed
• Each image must contain at least 2 objects

6
Outline

• Motivation Problem Statement
• Contributions
• Key Concepts
• Validation Methodologies
• Assumptions
• Rewrite today

7
Contributions

• Proposed a novel algorithm
• 9-Directional Spanning Area (9D-SPA)
• supports visualization, browsing, spatial
  reasoning, icon indexing and similarity retrieval
• Features of 9D-SPA
• Spatial reasoning and interpretation
• Visualization and database browsing
• A set of coarse-to-fine similarity for iconic
  indexing

8
Outline

• Motivation Problem Statement
• Contributions
• Key Concepts
• Validation Methodologies
• Assumptions
• Rewrite today

9
Key Concepts

• Indexing
• A method to efficiently and effectively store
  data to easily retrieve the information
• Spatial Operators
• Several relationships between two objects

10
Key Concepts

• Text-based vs. Content-based Image Search
• Text-based image annotations
• Example
• Google image search superman or super man
• Content-based image and the information within
  images
• Search image looks like above image in a database
• finger print matching
• suspect facial image matching,

11
Key ConceptsQuestion

• Which of the following queries are best suited
  for a text- or a content-based engine?
• Finger Print Matching?
• Thanksgiving related images?
• Suspect Facial Image Matching?

12
Key Concepts Related work

• 2D string (1987)Project an object along its x
  and y directions to form two strings representing
  the relative positions of objects in the x and
  y-axis.
• 9DLT 5 Topological relations (1997)
• 9 direction lower-triangular (9DLT) Disjoint,
  meet, partly overlap, contain and inside
• MOBR v.s. Centroid
• too sensitive to directional relations

13
Key ConceptsRelated work

• Applications of Content based image search system
• QBIC (Query By Image Content / IBM)
• Virage
• RetrievalWare
• VisualSEEK
• WaveGuide
• PhotoBook

14
Key Concepts Related work

• QBIChttp//www.hermitagemuseum.org/fcgi-bin/db2ww
  w/qbicSearch.mac/qbic?selLangEnglish

15
Key Concepts Related work

• QuerySEEKuser forms the queries by diagramming
  spatial arrangements of color regions.
  http//www.ctr.columbia.edu/jrsmith/html/pubs/acm
  mm96/acm.html

16
Key ConceptsContributions

• 9D-SPA
• 9 directional relations

17
Key Concepts Contributions

• 9D-SPA
• 5 topological relations
• 0 Disjoint
• 1 Meet
• 2 Partly-overlap
• 3 Cover
• 4 Contain or inside

18
Key Concepts Contributions

• Example
• TAB 0
• DAB
• DBA

19
Key Concepts Contributions

• Example
• Image visualization
• Similarity measure
• Directional similarity p2 to p1 is 0.67
• Topological similarity p2 to p1 is 1

20
Outline

• Motivation Problem Statement
• Contributions
• Key Concepts
• Validation Methodologies
• Assumptions
• Rewrite today

21
Validation Methodologies

• Experiments were performed on simulated synthetic
  data
• Advantages
• Able to measure the efficiency and effectiveness
  accurately of their algorithms
• Disadvantages
• The practicality of the system is unclear since
  no real data sets were used

22
Outline

• Motivation Problem Statement
• Contributions
• Key Concepts
• Validation Methodologies
• Assumptions
• Rewrite today

23
Assumptions

• Objects can be found beforehand through some form
  of image segmentation or object recognition
• Object recognition is a VERY hard problem
• Thus, those techniques should be clearly defined
  within the paper
• An exact definition of an object is not defined
• There may be a number of objects within an image
• A clear distinction between fore- and back-ground
  objects are not clear.

24
Outline

• Motivation Problem Statement
• Contributions
• Key Concepts
• Validation Methodologies
• Assumptions
• Rewrite today

25
Rewrite todayTechnical

• 5 Topological spatial relationships do not seem
  sufficient to image searches
• Since images use a 3D environment where
  topological images use a 2D environment
• Possible extensions of new operators in a 3D
  environment can consist of in front of or
  behind
• No comparisons between existing algorithms
• Apply to real world scenarios when a True Object
  Detection system exists

26
Rewrite todayStructural

• Claims need to be justified
• The most important feature of a symbolic picture
  is probably the binary spatial relationships
  between objects
• Explain image objects and icons and give general
  examples before start using them

27
Rewrite todayDiscussion

• Why not Z-curve or Herbert curve in 2D string?
• Why not combine MOBR and Centroid together in
  9DLT?

28
Questions?