Intrinsic Image Separation Using Weighted Map and Correction Using MRFs - PowerPoint PPT Presentation

1 / 26
About This Presentation
Title:

Intrinsic Image Separation Using Weighted Map and Correction Using MRFs

Description:

Intrinsic Image Separation Using Weighted Map and Correction Using MRFs. Robotics Lab ... A. A., Rangwala, S.& Hammmamji, 'Chromatic Properties of the Color Shading ... – PowerPoint PPT presentation

Number of Views:49
Avg rating:3.0/5.0
Slides: 27
Provided by: roboticsC
Category:

less

Transcript and Presenter's Notes

Title: Intrinsic Image Separation Using Weighted Map and Correction Using MRFs


1
Intrinsic Image Separation Using Weighted Map and
Correction Using MRFs
???????????????????
Robotics Lab Department of Computer Science and
Information Engineering National Cheng Kung
University
2
1. Introduction(1/2)
  • Motivation
  • Why separating Shading images and Reflectance
    images?
  • Reflectance images are more appropriate for
    pattern recognition, object detection and scene
    interpretation.
  • Shading images can be used for shading analysis,
    illumination assessment.

3
1.Introduction(2/2)
  • Image decomposition
  • A image can be decomposed into Shading and
    Reflectance images like

I(x,y) S(x,y) R(x,y)
H.G. Barrow and J.M. Tenenbaum, Recovering
Intrinsic Scene Characteristics from Images,
Computer Vision System, A. Hanson and E. Riseman,
eds., pp. 3-26. Academic Press, 1978.
4
Assumption
  • Our approach
  • Classify image derivatives
  • Each derivative is caused either by shading or
    reflectance ,but not both.
  • Derivatives caused by reflectance changes have a
    greater magnitude than those caused by shading.

Y. Weiss, Deriving Intrinsic Images from Image
Sequences, Proc. Intl Conf. Computer Vision,
2001.
5
2.System Flowchart
Module 2 Weighted-Map Creation and Derivative
Component Classification
Module 1 Intrinsic Derivative Component Creation
Input color Image
Classification Using Weighted Map
c
Module 3 Misclassification Correction Using
MRFs and Loopy Belief Propagation
Misclassification Correction
Module 4 Intrinsic Image Recovery
c
6
2.1 Module 1Intrinsic Derivative Component
Creation(1/2)
  • Logarithmic transformation

7
2.1 Module 1Intrinsic Derivative Component
Creation(2/2)
  • Derivative convolution

8
2.2 Intrinsic Derivative Component Creation
Classified result
9
2.3 Module 4Intrinsic Image Recovery Process
  • Deconvolution
  • Exponential transform
  • Composition

where
Y. Weiss, Deriving Intrinsic Images from Image
Sequences, Proc. Intl Conf. Computer Vision,
2001.
10
3.1 Module 2Part AColor Domain Transformation
  • LUM,RG,BY color space

Shading component
Shading componentonly in LUM image
plane!!
Reflectance componentin all three image
planes
Kingdom, F. A. A., Rangwala, S. Hammmamji,
Chromatic Properties of the Color Shading
Effect, Vision Research, 45, 1425-1437, 2005
11
3.2 Module 2 Part BFilter Convolution
12
3.3 Module 2 Part CWeighted-Map
Classification(1/2)
  • Reflectance-related map
  • Ideaextract reflectance component
  • Weighted-map

13
3.3 Module 2 Part CWeighted-Map
Classification(2/2)
  • Threshold classification

14
3.4 Experimental Results(1/2)
  • Intrinsic images

15
3.4 Experimental Results(2/2)
  • Intrinsic images

16
4. Misclassification
  • Problem
  • There are still some misclassifications after
    using weighted-map method.
  • Conclusion
  • Most derivatives on each edge are correctly
    classified as reflectance.
  • A small number of pixels on the same edge may be
    misclassified as shading.

17
4. Modeling Using Markov Random Fields(1/3)
  • Step1

where xi represents the hidden node state and yi
represents the observation node state at pixel i.


18
4. Modeling Using Markov Random Fields(2/3)
  • Step2 Initialize MRFs and define joint
    compatibility function.

1
1
1
1
0
0
1
1
1
1
1
0
1
1
0
0
1
1
1
0
1
0
0
1
1
1
0
0
1
1
19
4. Modeling Using Markov Random Fields(3/3)
  • Step 3Maximize objective function P by adjusting
    all hidden node states.

Original MRFs
Adjusting all hidden node states is time
consuming. Use Loopy Belief Propagation
to get a approximation solution.
MRFs after maximizing P
20
5. Experimental Results(1/3)
Misclassification Correction
No Misclassification Correction
21
5. Experimental Results(2/3)
22
5. Experimental Results (3/3)
Tappens Result
Our Result
Input image
Tappens Result
Our Result
Input image
M.F. Tappen, W.T. Freeman, and E.H. Adelson,
Recovering Intrinsic Images from a Single
Image, IEEE Trans. on Pattern Analysis and
Machine Intelligence, Vol. 27, No. 9, pp.
1459-1472, 2005.
23
Thanks for your attention
24
Appendix
  • J.S. Yedidia, W.T. Freeman, and Y. Weiss,
    Understanding Belief Propagation and its
    Generalizations, MITSUBISHI Electric Research
    Lab, TR-2001-22, 2002

25
1.Introduction(1/3)
  • The Goal
  • A image is composed of two parts, called Shading
    and Reflectance images. We proposed a method for
    separating Shading and Reflectance images given a
    single input image.
  • Definition
  • What are Shading and Reflectance?
  • Reflectance Remain constant under different
    illumination conditions.
  • Shading Vary from different illumination
    conditions.

25
H.G. Barrow and J.M. Tenenbaum, Recovering
Intrinsic Scene Characteristics from Images,
Computer Vision System, A. Hanson and E. Riseman,
eds., pp. 3-26. Academic Press, 1978.
26
3. Weighted-Map Method Flowchart
Part A Color Domain Transformation
Part B Filter Convolution
Part C Weighted-Map Classification
26
Write a Comment
User Comments (0)
About PowerShow.com