A different approach to fitting transformations in the presence of outliers is to model the uncertainty as a mixture of two Gaussians. The first Gaussian models the image noise, and the sec- ond Gaussian, which has a very large variance, accounts for the outliers. For example, for the affine transformation we would have
Pr(x|w)=λNormxfaff[w,Φ,τ],σ2I+(1−λ)Normxfaff[w,Φ,τ],σ2
0I+
whereλis the probability of being an inlier,σ2is the image noise, andσ2is the large variance that accounts for the outliers. Sketch an approach to learning the parametersσ2,Φ,τ, andλof this model. You may assume thatσ2is fixed. Identify a possible weakness of this model.
0
Already registered? Login
Not Account? Sign up
Enter your email address to reset your password
Back to Login? Click here