Title: Evaluation of processes used in screen imperfection algorithms
1Evaluation of processes used in screen
imperfection algorithms
2Introduction
- Screen compensation algorithm
- Divided in four parts
- Projector characterization
- Camera characterization
- Geometrical alignment
- Screen compensation
- A Projection System with Radiometric
compensation for Screen Imperfections, Nayar et
al. - Making One Object Look Like Another Controlling
Appearance Using a Projector-Camera System,
Grossberg et al. - Robust Content-Dependent Photometric Projector
Compensation, Ashdown et al.
3Motivation
- Screens increases the cost of projectors
- Screens takes up space
- Screens decreases projectors mobility
- And therefore decreases functionality.
- Can alter color of objects (Virtual offices).
4Index
- Thesis
- General
- Goal
- General model for characterization
- Projector
- Camera
- Geometrical alignment
5Thesis-general
- This thesis focus on the different steps of
achieving screen independence. - Evaluated 2 projector characterization methods
and established their parameters. - Evaluated 4 camera characterization methods and
established their parameters. - Transformation of coordinates of the screen from
the captured image to the original image. - Use of regression to compensate for the screens
effect.
6Thesis- general
Colors are modified by the projector.
Color I is projected
Colors are modified by the screen
Camera captures projected colors. Colors are
again modified, this time by the camera
7Thesis - general
- Input and output devices are restricted by their
sensors and/or ability to reproduce colors. - To be able to calculate how screens modify
colors, we need to know how input and output
devices modify them first.
8Thesis-Goal
- Evaluate characterization methods for camera
- Evaluate characterization methods for projectors
- Implement Geometrical alignment algorithm
- Investigate the effect of screen compensation as
the characterization error changes.
9General model of characterization
Ex.Spline interpolation
10Projector Resarch Questions
- How many colors are needed for linearization
using linear, spline and cubic interpolation? - How will PLCC compare against a characterization
using regression? - How many colors in the training set is needed to
for the color difference to be considered hardly
visible, when regression is used?
11Projector - Characterization methods
- 3 different interpolation techniques for
linearization. - Piecewise Linear assuming constant chromaticity
model (PLCC). - Regression
12Projector-experiment
Gamut of the projector
Color difference is calculated for different
amount of colors used in linearization and as
trainining-set. PLCC do no require
training-set. Different interpolaiton techniques
was used to linearize RGB.
51 colors for the training-set
33 colors pr ramp
150 Random colors
100 colors for test-set
10 to 20 colors
10 to 20 colors
13Projector conclusion
- PLCC performed better than regression. With only
12 colors used in linearization acceptable result
is achieved. - Possible threat The assumptions of the PLCC
model is correct for the test-set but not for the
whole gamut. - It is possible to achieve good result with
regression using 12 or more colors for
linearization and 12-18 colors in the
training-set.
14Camera Research questions
- How many colors should be used for regression?
- What order of polynomial regression should we
use? - How will the use of only the cubic root function
before transformation to LAB perform? - How will use of CIELAB compare to CIEXYZ?
- Will always the method that performs best in
CIEXYZ perform best also in CIELAB? - How stabile are these methods?
15Camera characterization methods
Method name Method description
Method 1 Gamma method for linearization and regression into CIEXYZ space
Method 2 Polynomial fitting for linearization and regression into CIEXYZ space
Method 3 No linearization beyond a cubic root function and regression into CIELAB space
Method 4 Gamma method and a cubic root function for linearization and regression into CIELAB space
Method 5 Polynomial fitting and a cubic root function for linearization and regression not CIELAB space
16Camera Experiment
- Regression up to fourth order was used.
- Methods were tested 100 timer per training-set.
- 180 random colors were measured
- 33 grey values were used for linearization.
17Camera-Result
Size of regression Matrix Method 1 Method 2 Method 3 Method 4 Method 5
3x3 10.35 7.77 19.66 9.03 7.80
3x5 8.11 7.18 16.29 8.21 6.18
3x10 6.20 3.97 6.58 3.51 3.75
3x20 4.52 2.24 2.82 1.79 2.53
3x35 3.20 1.40 1.34 1.10 1.37
18Camere-conclusion
- Number of colors used for regression was
dependent on methods and order of regression. - Minimum order Second order regression.
- Use of cubic root function proved to yield good
results but was very unstabile. - CIELAB performed better than CIEXYZ and was more
stabile. - Its not certain that method that perfoms well in
CIEXYZ performs as well in CIELAB. (Method 1 and
4 versus Method 2 and 5). - Stability was dependent on amount of colors in
the training-set, order of regression and
linearization method.
19Geometrical alignment.
20Geometrical alignment
- The points are detected
- Each point are binary coded.
- Divided in blocks
- Regression for finding transformation matrix.
- Compensation
- Divide image in blocks.
- Multiply with the transformation matrix.
- Dependent on size of the screen, the resolution
of the camera and number of points and blocks.
21Acknowledgement
- I want to thank Mr. Hardeberg and HiG
administration for giving me chance to visit
Japan. - I want also to thank Tsukdada-san, Toda-san,
Funyama-san, Inoue-san and rest of the NEC
employees who have welcomed me warmly.
22Resten av slides er bare i tilfelle jeg trenger
dem.
23ProjectorMean Delta
24ProjectorMean Delta
25Projector interpolationregression
26ProjectorInterpolationregression
27Camera-standard deviance.