The result is low ?

[chenkang455]

Hi @zwx8981 @guspan-tanadi , i use the following code to test the LIQE metric, though the img is sharp, the liqe result seems relatively low, can you be kindly explain it?

import torch
import cv2
import torch.nn.functional as F
import pyiqa
iqa_metric = pyiqa.create_metric("liqe_mix", device="cuda")
img = cv2.imread('exp/REDS_cs_final/final_iter10/imgs/010_0010_sharp.png')
img = torch.tensor(img)[:,:,0]
img = img[None,None].cuda() / 255
short_edge = 384
h,w = img.shape[2],img.shape[3]
if h < w:
    new_h, new_w = short_edge, int(w * short_edge / h)
else:
    new_h, new_w = int(h * short_edge / w), short_edge
img = F.interpolate(img, size=(new_h, new_w), mode='bilinear', align_corners=False)
print(img.shape)
print(iqa_metric(img))
del img

torch.Size([1, 1, 384, 614])
tensor([1.5381], device='cuda:0')

[zwx8981]

Hi, I think it is partially because this is a grayscale image. While LIQE is exposed to RGB images during training, there is some domain gap.