Project DetailsNowadays, the metric used to calculate the statistic distance between different datasets is the FID (Frchet Inception Distance): it uses a pretrained inception network and a divergence very close to the W2 divergence (in this case) to approach the distance between them. We assume that the latent representation of each dataset has a Gaussian distribution. We also assume that the Gaussian distribution is not degenerate: we assume that the covariance matrix is a positive definite matrix: all the principal components are not zeros. To these assumptions, we can add the numerical instability and the impossibility to score a single image.
The limits of this method had made us need to look for a new metric. Can we find a new solution that wont need those assumptions and with whom we can score a single image while maintaining numerical stability?
