However in the paper (A Perceptually Motivated Online Benchmark for Image Matting) where these metrics are first defined, these error metrics are defined like this with the best parameters:
G(A, Agt) = sum((∇αi − ∇αi∗)^2) for all pixels i
C(A,Agt) = sum((φ(αi , Ω) − φ(αi∗ , Ω))) for all pixels i
but I am not sure if this paper follows the same metric (instead of summing, they are taking the mean), or uses L1 or Euclidean norm. At first, I assume they are using MSE as both connectivity and gradient error metric, but I'm not sure. Can anyone help me on that ?
In the paper, we can see that connectivity and gradient errors are defined as
G(A, Agt) = 1/K || ∇Ai − ∇Ai_gt || C(A,Agt)= 1/K || φ(Ai,Ω) − φ(Ai_gt,Ω) ||,
However in the paper (A Perceptually Motivated Online Benchmark for Image Matting) where these metrics are first defined, these error metrics are defined like this with the best parameters:
G(A, Agt) = sum((∇αi − ∇αi∗)^2) for all pixels i C(A,Agt) = sum((φ(αi , Ω) − φ(αi∗ , Ω))) for all pixels i
but I am not sure if this paper follows the same metric (instead of summing, they are taking the mean), or uses L1 or Euclidean norm. At first, I assume they are using MSE as both connectivity and gradient error metric, but I'm not sure. Can anyone help me on that ?