| | Introduction | Details |
|---|
| FID ↓ | FID [ 28 ] is robust to noise and consistent with human perception. It uses the Fréchet distance between two multivariate Gaussians FID(Pr,Pg)=∥μr−μg∥2+Tr(Cr+Cg−2(CrCg)21). | Pr , Pg indicate the distribution of real samples and generated samples respectively. C is the covariance matrix of feature vector. Tr is the trace of matrix. We follow [ 28 ] to use pre-trained InceptionV3 [ 29 ] network to extract features. |
|---|
| KID ↓ | KID [ 30 ] can be seen as an improved version of FID [ 28 ] , its estimates are unbiased and asymptotically normal KID(Pr,Pg)= Exr,xr′∼Pr,xg,xg′∼Pg[k(I(xr),I(xr′)) −2k(I(xr),I(xg))+k(I(xg),I(xg′))]. | I(xr) , I(xg) are features obtained from real samples and generated samples respectively. k indicates the polynomial kernel function, k(a,b)=(d1aTb+1)3 , where d is the representation dimension. We follow [ 30 ] to use pre-trained InceptionV3 [ 29 ] network to extract features. |
|---|
| KMMD ↓ | KMMD [ 31 ] uses fixed kernel function to measure the dissimilarity between real samples and generated samples MMD2(Pr,Pg)= Exr,xr′∼Pr,xg,xg′∼Pg[k(xr,xr′) −2k(xr,xg)+k(xg,xg′)]. | xr and xg is the sample feature from true distribution and generated distribution. k is the gaussian kernel function. Results in [ 32 ] show KMMD appears to be a good metric in terms of discriminability, robustness and efficiency with pre-trained ResNet [ 33 ] . Thus, we use pre-trained ResNet [ 33 ] for feature extraction. |
|---|
| IS ↑ | IS [ 34 ] has a reasonable correlation with human judgment of image quality and is able to show the diversity of generated images IS(Pg)= exp(Ex∼Pg[KL(pM(y∣x)∣∣pM(y))]). | pM(y∣x) is the label distribution of x as predicted by model M , and pM(y)=∫xpM(y∣x)dPg . KL indicates Kullback-Leibler divergence. We follow [ 34 ] to use InceptionV3 [ 29 ] network to extract the probability distribution of each category. |
|---|
| PSNR ↑ | PSNR is commonly used to measure the quality of image by comparing samples with original images pixel-wisely PSNR= 10×log10(MSEMAXI2). | MAXI is the maximum possible pixel value of the image. MSE is the mean-square error. |
|---|