iterative and interactive search for objects

22
Iterative and Interactive Search for Objects Moty Golan & Oren Kerem Instructor: Dr. Sigal Ar

Upload: rod

Post on 15-Jan-2016

32 views

Category:

Documents


0 download

DESCRIPTION

Iterative and Interactive Search for Objects. Moty Golan & Oren Kerem Instructor: Dr. Sigal Ar. Choose object. Search. Mark Results. Introduction. Databases of objects choose one working database choose object from the database arrange database by similarity to chosen object - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Iterative and Interactive  Search for Objects

Iterative and Interactive Search for Objects

Moty Golan & Oren Kerem

Instructor: Dr. Sigal Ar

Page 2: Iterative and Interactive  Search for Objects

Introduction

• Databases of objects– choose one working database– choose object from the database– arrange database by similarity to chosen object

• Interactive: results can be marked as “good” or “bad”

• Iterative: marked results used to refine the search

Search

Chooseobject

MarkResults

Page 3: Iterative and Interactive  Search for Objects

Databases

• 3D objects’ surface– 10000 sampling points (no color)– Uniform distribution– Location of point and normal to the surface

• 2D color pictures– Distinct families of pictures– HSV color representation

Page 4: Iterative and Interactive  Search for Objects

Objects’ Metric

• Object representation - feature vector

• Initial distance between objects - Euclidean distance: d(Dx , Dy) = || X - Y ||2

• Iterations: weighted Euclidean distance

Page 5: Iterative and Interactive  Search for Objects

• Weights

• Weighted distance : d(Dx , Dy) = [X - Y ]T W [X - Y] + b

Page 6: Iterative and Interactive  Search for Objects

Feature Vectors

• Data of a certain type is extracted from the object

• A vector of values is calculated based on the data

• Feature type – the method of the values calculation– Moments– Histograms

Page 7: Iterative and Interactive  Search for Objects

Data Types

• One aspect of the object is represented by a finite set of values

• Naturally numerous aspects exist• For graphic object use visual aspects:

– Pixels– Normals– Curvature– Color (HSV representation)

Page 8: Iterative and Interactive  Search for Objects

Implemented Data and Feature Types

3D

Moments Histograms

points normals points normals2D

Moments Histograms

curvature HSV points curvature HSV points

Page 9: Iterative and Interactive  Search for Objects

Goals

• Organize the databases

• Design an interface to enable searches on both databases using all parameters

• Conduct searches to explore behavior

Page 10: Iterative and Interactive  Search for Objects

Organizing the Databases

• Extracting the data types

• Calculation of feature vectors

• Create object icons

• Build directory hierarchy

Read_Graphic_Body Create_Icon

Compute_Moments

Organize_3dBase_Mom

Extract_MomFeatures

3D Example:

Page 11: Iterative and Interactive  Search for Objects

Calculating Features

• Normalize the data if needed – Ex. according to 1st & 2nd order moments

• Calculate moment vectors according to desired orders

• Calculate and flatten histograms

• Save in appropriate files

Page 12: Iterative and Interactive  Search for Objects

Icons

• “Flatten” and scale the 3D objects

• Sub-sample the 2D objects

Page 13: Iterative and Interactive  Search for Objects

Organized Databases

2d 3d sample

3d

images

2d

features

root

3d2d 3d2d

Page 14: Iterative and Interactive  Search for Objects

Interface Options

• Select database: 2D / 3D• Choose object• Set search parameters

– Feature kind and level

– Data type

• Conduct search• Database viewing & browsing

– Full data display

• Good and bad object marking

Page 15: Iterative and Interactive  Search for Objects
Page 16: Iterative and Interactive  Search for Objects

Program FlowDiagram

STARTING POINT

CHOOSE OBJECT

WAITING TO

SEARCH

WAITING FOR

MARKS

MARKING OBJECTS

Choose Object Button

Choose This Button

Search Button

Mark Results ButtonUpdate Search Button

Choose Object ButtonANYWHERE

Change search parameters

/ Change Database

Page 17: Iterative and Interactive  Search for Objects

Implementation: State Machine

• Each program stage is a state:– Program waiting stages– Program action stages

• User actions switch between states

• In each waiting state the relevant options are made available

while

switch

case a: ... case b: ... . . .

end

end

Page 18: Iterative and Interactive  Search for Objects

Cases Diagram

case 1

case 3

case 5case 23

case 25

30

10

100

Choose This pressed

Search pressed

50

case 2Mark Results

pressed

Update Search pressed

ANYWHERE

Good Object pressed

Bad Object pressed

Change Object / Database

INITIAL CASE

command case

user action

Legend

empty

Change Parameters

Page 19: Iterative and Interactive  Search for Objects

Demonstration

Page 20: Iterative and Interactive  Search for Objects

Test Runs

• Good results– Most tests converged within 3-4 iterations– Increase in “good” objects in each iteration– “Bad” objects removed from top 30 results

• Limitations– Database variety– Computation demands vs. response time– Human perception vs. object representation

Page 21: Iterative and Interactive  Search for Objects

Feature Explorations

• Higher feature levels do not guarantee better results

• Completeness of representation does not guarantee better results

• Some tests affected by feature type

Page 22: Iterative and Interactive  Search for Objects

Conclusions

• Database arrangement and interface were implemented as shown

• Various searches show that the algorithm successfully identifies classes of objects

• Further research can include:– The effect of specific high feature levels– The effect of computation constants– Combining different feature types and levels