Title: Intrinsic Image Separation Using Weighted Map and Correction Using MRFs
1Intrinsic Image Separation Using Weighted Map and
Correction Using MRFs
???????????????????
Robotics Lab Department of Computer Science and
Information Engineering National Cheng Kung
University
21. 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.
31.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.
4Assumption
- 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.
52.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
62.1 Module 1Intrinsic Derivative Component
Creation(1/2)
- Logarithmic transformation
72.1 Module 1Intrinsic Derivative Component
Creation(2/2)
82.2 Intrinsic Derivative Component Creation
Classified result
92.3 Module 4Intrinsic Image Recovery Process
- Deconvolution
- Exponential transform
- Composition
where
Y. Weiss, Deriving Intrinsic Images from Image
Sequences, Proc. Intl Conf. Computer Vision,
2001.
103.1 Module 2Part AColor Domain Transformation
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
113.2 Module 2 Part BFilter Convolution
123.3 Module 2 Part CWeighted-Map
Classification(1/2)
- Reflectance-related map
- Ideaextract reflectance component
- Weighted-map
133.3 Module 2 Part CWeighted-Map
Classification(2/2)
143.4 Experimental Results(1/2)
153.4 Experimental Results(2/2)
164. 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.
174. Modeling Using Markov Random Fields(1/3)
where xi represents the hidden node state and yi
represents the observation node state at pixel i.
184. 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
194. 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
205. Experimental Results(1/3)
Misclassification Correction
No Misclassification Correction
215. Experimental Results(2/3)
225. 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.
23Thanks for your attention
24Appendix
- J.S. Yedidia, W.T. Freeman, and Y. Weiss,
Understanding Belief Propagation and its
Generalizations, MITSUBISHI Electric Research
Lab, TR-2001-22, 2002
251.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.
263. Weighted-Map Method Flowchart
Part A Color Domain Transformation
Part B Filter Convolution
Part C Weighted-Map Classification
26