FiganLoss<T>

Helpers & Utilities · AiDotNet.Video.FrameInterpolation

FIGAN’s multi-scale perceptual objective: an L1 + VGG content loss evaluated at every scale, and combined with a very small adversarial term.

For Beginners

The model is graded on how close its invented frame is to the real one, measured both pixel by pixel and through the eyes of a pretrained image network, at three levels of detail at once — plus a small nudge from a critic trying to spot fakes.

How It Works

van Amersfoort et al., “Frame Interpolation with Multi-Scale Deep Loss Functions and Generative Adversarial Networks” (arXiv:1711.06045).

The coefficients are specified by the paper, not free parameters.

Three details are easy to get wrong and all three matter:

Methods

Method Summary
double Content(Vector<T> predicted, Vector<T> target, Vector<T> predictedFeatures, Vector<T> targetFeatures) Eq.
double DiscriminatorObjective(double discriminatorOnReal, double discriminatorOnFake) The discriminator’s objective, log D(real) + log(1 - D(fake)) (Eq.
double GeneratorAdversarial(double discriminatorOnFake) The generator’s adversarial term, log(1 - D(G(I_0, I_1))) (Eq.
double Total(double finestScale, IReadOnlyList<double> coarserScales, double refinedSynthesis, double adversarial) Eq.

Fields

Field Summary
double AdversarialWeight The paper’s adversarial weight in the total loss (Eq.
double CoarseScaleWeight The paper’s weight on the two COARSER synthesis scales (Eq.
double VggWeight The paper’s VGG weight in the content loss (Eq.