Abstract
Modeling Color Changes Between Image matching has many important
applications, including 3D reconstruction and
Cameras Using a Large Database of robot navigation. If images are taken by
different cameras or at different times of day, the
colors can change considerably. The goal of
Registered Images this work was to investigate ways to model such
color changes. We created a test database of
registered images while changing camera
model, illumination, viewing angle, white
balance, and exposure. We investigated
various image registration techniques and wrote
a script to align large batches of images using
the GDB-ICP algorithm. We created six
datasets consisting of over 1200 images. Using
Kelvin Gorekore Scott Wehrwein Daniel Scharstein Amy Briggs C++ and Matlab, we built several tools to
experiment with different color models. We
Funding for this project was provided by the National Science Foundation under grants IIS-0713442 and IIS-0413169 used two-dimensional histograms to visualize
color changes and investigated different curve-
fitting techniques to model the histograms. Our
results show that the color changes are smooth
Motivation Goals but nonlinear, and that color channels cannot be
While color should be helpful for • Create a large dataset of images taken by different cameras while modeled separately. Realistic color models will
become increasingly important for matching
image matching, different changing viewpoint, lighting, white balance, and exposure images from online photo collections; we hope
cameras and lighting conditions that our database and initial findings will be
introduce complex color changes • Accurately model color changes between the images valuable to future research in this area.
Creating Datasets Visualizing Color Changes
Requirements • Pixel-by-pixel comparison of each color channel using 2D histogram
• Generated web pages to display comparisons among a set of images:
• Different cameras • Different camera settings: white
• Different shooting conditions: balance and exposure
viewpoints and illuminations • Same scene, pixel-accurate
alignment
6 Scenes:
255
Image 2 intensity
0
0 255
Image 1 intensity
Overview table
Details of a single comparison
Biwall – 45 Images Chalk – 98 Images Modeling Color Changes
• R1, G1 and B1 are red, green and blue color channels of the first image
• R2, G2 and B2 are red, green and blue color channels of the second image
• Modeling color channels independently:
R2 = fR(R1), G2 = fG(G1) and B2 = fB(B1)
• Models considered: - linear: f(x) = ax + b
- scaled exp: f(x) = sxg
- polynomial: f(x) = akxk + ak-1xk-1+……+a0
Tiles – 48 Images Poster – 72 Images
• Modeling color twist:
R 2 A11 A12 A13 R1 More generally:
G 2 A21 A22 A23 G1 R2 = fR(R1, G1, B1),
B = f (R , G , B )
G2 = fG(R1, G1, B1),
B 2 A31 A32 A33 B1 for smooth functions f, e.g quadratic
2 B 1 1 1
Quadratic color twist,
Scaled exponential: Polynomial: smoothing, and polynomial:
Podium – 70 Images Color – 222 Images, each cropped into
four sections , plus 72 raw/jpeg pairs
5 Cameras: Canon EOS 20D, Canon Powershot G1, Sony A100, Image 1 Corrected Image 1 Image 1 Corrected Image 1 Image 1 Corrected Image 1
Canon Powershot A540, Casio Elilim EX-S10
Image Alignment Image 2 Residual Image 2 Residual Image 2 Residual
• Used Generalized Dual-Bootstrap Iterative Closest Point algorithm for
alignment
• Created batch scripts to align entire sets of images, gracefully handled
alignment failures
Before alignment After alignment Overlay of aligned images
RGB Histogram RGB Histogram RGB Histogram
Conclusions
• Color changes are smooth but nonlinear
• Channels cannot be modeled independently for images from
different cameras