lightfield& montages0vis.berkeley.edu/.../images/1/1e/vdfinalpresentation.pdfmotivation why use...
TRANSCRIPT
Light Field Montages
Vera Dadok
Problem Given multiple light field shots of the a scene with minor
differences, how can you successfully combine the best parts of each image?
Challenges: Many representations of light fields (format) Large file sizes, higher dimensionality Creation and display of light fields
Image Source: Light Field Rendering, 1996
Motivation Why use Light Fields? • Hand-held
plenoptic cameras • Light Fields contain
more information than images (4-d vs. 2-d sample)
• Few tools to modify light fields exist, and many interesting operations to develop Image Source: Lytro.com
Motivation Why montage Light Fields? • Single Light Field may not
capture ideal scene o Want to remove/change elements
• Automatic montage method necessary o Large image stacks or
representations
Refined Problem Statement
• Input o Two Light Fields of the same scene with minor differences represented as
an array of images taken at different (u,v) coordinates
o User selection of pixels fixed to each source
• Output o Light Field Montage containing the best of the two images
Image Source: Light Field Rendering, 1996
Refined Problem Statement
• Input o Two Light Fields of the same scene with minor differences represented as
an array of images taken at different (u,v) coordinates
o User selection of pixels fixed to each source
• Output o Light Field Montage containing the best of the two images
Approach
• Ingredients: o Light Fields o Light Field Viewer o User Interaction Algorithm o Montage Algorithm
Input: 2 Light Fields
User interaction: Select pixels from
sources
Montage Algorithm
Result: Light Field Montage
User Interaction • Light field is represented as a matrix of images,
each positioned at a (u,v) coordinate. • User interacts only with the 4 corners
V-‐‑position
U-‐‑position
V-‐‑position
U-‐‑position
Screenshot
Light field montage Based on Graph Cut (2d and 3d) • Minimize a Cost Function (Energy Function)
1
2
o Data Term • Based on user-selected pixels and
surrounding pixels o Augmented user-selection with inverse
square law based on distance • Interpolated pixels at (u,v) coordinates from
corner pixel selections
o Smoothness Term • Several possible functions: Color, Color &
Gradient, Gradient, Color & Edges (see Agarwala et. al, SIGGRAPH 2004)
• Color cost – related to difference in pixel colors between 2 source light fields
Image: UC Berkeley CS294-‐‑69 Course Notes, 2011
min E = min (Data Term + Smoothness Term)
Graph-‐‑cut at work (2-‐‑d)
3d Graph Cut
Graph cut on ordered stack of images from single line of u or v
Results
Demos
Plant Demos: Originals: file:///Users/Vera/Desktop/apertureupdateb/aperture.swf?lightfield=cdswithoutplantscroppedsmaller&xmlzip=cdswithplants.zip file:///Users/Vera/Desktop/apertureupdateb/aperture.swf?lightfield=cdswithplantscroppedsmaller&xmlzip=cdswithplants.zip
Graph Cuts: file:///Users/Vera/Desktop/apertureupdateb/aperture.swf?lightfield=plants2d&xmlzip=cdswithplants.zip file:///Users/Vera/Desktop/apertureupdateb/aperture.swf?lightfield=plants3d&xmlzip=cdswithplants.zip
Jellybean Demo: Originals: file:///Users/Vera/Desktop/apertureupdateb/aperture.swf?lightfield=preview&xmlzip=previewc.zip file:///Users/Vera/Desktop/apertureupdateb/aperture.swf?lightfield=newColoredBeans2&xmlzip=previewc.zip
Graph Cut: file:///Users/Vera/Desktop/apertureupdateb/aperture.swf?lightfield=previewnewColoredBeans2montage&xmlzip=previewc.zip
Future Work • 4d graph-cut • Improved user interactions
o What is the most intuitive way to edit a light field? o Automatic reality checks using known geometry/depth o Speed o Multiple light fields montage (just layer via user interface) o Additional montage modes (lighting, etc.)
• New cost functions for graph cut o Use depth maps, geometry
• Obtain more light fields to test this on o Or create synthetic light fields.
• Manipulating different representations of light fields o Depth array of images (full aperture)
Thank You!
Selected references • Images :
o UC Berkeley CS294-69 Course Notes, 2011 o Lytro.com o Light Field Rendering, Marc Levoy and Pat Hanrahan, 1996.
• Light Fields: o Stanford Repository, http://lightfield.stanford.edu/, December 2011
• Code supplements: o Graphcut:, Matlab wrapper for robust higher order potentials. Shai Bagon,
http://www.wisdom.weizmann.ac.il/~bagon/matlab.html 2009. • See also: http://www.wisdom.weizmann.ac.il/~bagon/matlab_code/robustpn_wrapper.bib
o Light Field Renderer: From http://lightfield.stanford.edu/, o FZip
• Code supplement references: o Efficient Approximate Energy Minimization via Graph Cuts. Yuri Boykov, Olga Veksler, Ramin Zabih,
IEEE transactions on PAMI, vol. 20, no. 12, p. 1222-1239, November o What Energy Functions can be Minimized via Graph Cuts? Vladimir Kolmogorov and Ramin Zabih.
European Conference on Computer Vision (ECCV), May 2002. o An Experimental Comparison of Min-Cut/Max-Flow Algorithms for Energy Minimization in Vision. Yuri
Boykov and Vladimir Kolmogorov. In IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), September 2004
For a more thorough list of references, please contact Vera Dadok