GAN metrics - Serbipunk/notes GitHub Wiki
cross - reenactment
self - reenactment
computable scores
PSNR (Peak Signal-to-Noise Ratio)
SSIM (Structural Similarity Index Measure)
from skimage.metrics import structural_similarity
theory: https://www.youtube.com/watch?v=-i3NQ-by2b8
LPIPS (Learned Perceptual Image Patch Similarity)
orb_sim
def orb_sim(im1, im2):
orb = cv2.ORB_create()
kp_a, desc_a = orb.detectAndCompute(im1, None)
kp_b, desc_b = orb.detectAndCompute(im2, None)
# bf matcher
bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
# perform matches
matches = bf.match(desc_a, desc_b)
similar_regions = [m for m in matches if m.distance < 50]
if len(matches) == 0:
return 0.
return len(similar_regions) / len(matches)
FID (Frechet Inception Distance)
https://www.youtube.com/watch?v=TJeeZFNXi9M
IQA (image quality assessment metrics)
CSIM (just for face generation)
cosine similarity of fr embedding
user preference scores
UMTN & UAPP (megaportraits)