Should images be normailzed to [-1, 1] or [0, 1]?

[luccachiang]

Hi, thanks for your codebase! In calculate_lpips.py, it says image should be RGB, IMPORTANT: normalized to [-1,1]. https://github.com/JunyaoHu/common_metrics_on_video_quality/blob/c12e449540cf27c1b1412c8012ecc35413fda6a7/calculate_lpips.py#L26C59-L26C59 However, in the README notice section, it says Make sure the pixel value of videos should be in [0, 1].

common_metrics_on_video_quality/README.md

Line 167 in c12e449

3. Make sure the pixel value of videos should be in [0, 1].

. How to normalize to raw video whose values is in range [0, 255]? Would you please provide detailed value ranges of all the four metrics?

[JunyaoHu]

Hello, I am working for CVPR deadline, please wait for my answer.

[luccachiang]

Hello, I am working for CVPR deadline, please wait for my answer.

Same as well... Wish you good luck.

[JunyaoHu]

the inital video input range [0-1].

FVD -> [0-1] to [0-255]

common_metrics_on_video_quality/fvd/videogpt/fvd.py

Lines 51 to 55 in c12e449

	def preprocess(videos, target_resolution=224):
	# we should tras videos in [0-1] [b c t h w] as th.float
	# -> videos in {0, ..., 255} [b t h w c] as np.uint8 array
	videos = einops.rearrange(videos, 'b c t h w -> b t h w c')
	videos = (videos*255).numpy().astype(np.uint8)

SSIM -> no trans [0-1] to [0-1]

common_metrics_on_video_quality/calculate_ssim.py

Lines 26 to 27 in c12e449

	def calculate_ssim_function(img1, img2):
	# [0,1]

LPIPS -> [0-1] to [-1-1]

common_metrics_on_video_quality/calculate_lpips.py

Lines 15 to 36 in c12e449

	def trans(x):
	# if greyscale images add channel
	if x.shape[-3] == 1:
	x = x.repeat(1, 1, 3, 1, 1)
	# value range [0, 1] -> [-1, 1]
	x = x * 2 - 1
	return x
	def calculate_lpips(videos1, videos2, device):
	# image should be RGB, IMPORTANT: normalized to [-1,1]
	print("calculate_lpips...")
	assert videos1.shape == videos2.shape
	# videos [batch_size, timestamps, channel, h, w]
	# support grayscale input, if grayscale -> channel*3
	# value range [0, 1] -> [-1, 1]
	videos1 = trans(videos1)
	videos2 = trans(videos2)

PSNR -> no trans [0-1] to [0-1]

common_metrics_on_video_quality/calculate_psnr.py

Lines 6 to 8 in c12e449

	def img_psnr(img1, img2):
	# [0,1]
	# compute mse

---

How to normalize ?

if you input is [0-255]

you can

video = video/255.0

and then calculate

---

Wish you good luck too.

---

image should be RGB, IMPORTANT: normalized to [-1,1] is copied form other repos like