effective optical flow estimation jan kamenický21.10.2011
TRANSCRIPT
Effective Optical Flow Estimation
Jan Kamenický 21.10.2011
Motivation
Usage
• Motion detection• Object segmentation• Video encoding (compression)• Stereo disparity measurement
Optical flow equation
• Color constancy
• Taylor series
• Optical flow equation
Estimating optical flow
• Basic equation
• Local methods– Lucas & Kanade• flow field is locally constant (or affine)• least squares minimization• cannot handle interior parts of objects
Estimating optical flow
• Basic equation
• Global methods– Horn & Schunk
• more sensitive to noisedata term smoothing term
• Data term– non-homogenous surface (shading, reflections)– non-flat scene / non-uniform lighting– spatial discontinuities
• Smoothness term– discontinuities (moving objects boundaries)
Problems
Estimating partial derivatives
• Discrete approximation by differences– forward, backward – not exact– use 2x2x2 cube in (x,y,t) space• compute the difference as an average of 4 adjacent first
order differences
– use larger support• e.g. [1, -8, 0, 8, -1]/12
i i+1
j
j+1
kk+1
Data term
• L2 norm
• L1 norm
• Many modifications– generalized Charbonnier
Regularization term
• Enforces smooth flow field• Similar norms can be used– L2, L1 (total variation), …
• Other possibilities– Laplacian instead of gradient
Dealing with larger displacements
• Smoothing (blurring)– usually Gaussian kernel– decreases flow field accuracy
• Pyramidal approach– compute flow on down-sampled images– up-sample the flow to next level– compute the warping (using the optical flow)– repeat
More optimizations
• Graduated non-convexity– iteratively move from convex energy function to
the more robust non-convex form
• Median filtering (5x5)– weighted modification
• More warping steps on one pyramid level
OF methods comparison
• Optical flow estimation benchmark– http://vision.middlebury.edu/flow/
• Average end-point error
References
• Main described method– D. Sun, S. Roth, M. J. Black: Secrets of Optical Flow
Estimation and Their Principles, CVPR 2010– http://www.cs.brown.edu/~black/code.html