cutcutcodec.core.analysis.video.metric

Image metrics.

Functions

`compare`(ref, dis, **kwargs)	Compare 2 video files with differents metrics.
`lpips`(ref, dis, args, *kwargs)	Compute the Learned Perceptual Image Patch Similarity.
`psnr`(ref, dis, args, *kwargs)	Compute the peak signal to noise ratio of 2 images.
`ssim`(ref, dis, *args[, stride])	Compute the Structural similarity index measure of 2 images.
`uvq`(dis[, _model])	Compute the Perceptual Video Quality.

Details

cutcutcodec.core.analysis.video.metric.compare(ref: Path | str | bytes, dis: Path | str | bytes, **kwargs) → dict[str, list[float]][source]

Compare 2 video files with differents metrics.

Parameters

refpathlike: The reference video file.
dispathlike: The distorted video.
lpips_alexboolean, default=False: If True, compute the lpips with alex (medium).
lpips_vggboolean, default=False: If True, compute the lpips with vgg (slow).
psnrboolean, dafault=False: If True, compute the psnr (very fast).
ssimboolean, default=False: If True, compute the ssim (slow).
uvqboolean, default=False: If True, compute the uvq on the dis video (very slow). It returns only one value per second. If you want to compute this metric only, give None to ref.
vmafboolean, default=False: If True, compute the vmaf (medium).

Returns

metricsdict[str, list[float]]: Each metric name is associated with the scalar value of each frame. All the numbers are rounded to 4 decimals number.

Notes

Frames are converted to yuv if not already converted, then the distorted video is converted to the color space of the reference video.

Examples

>>> import pprint
>>> from cutcutcodec.core.analysis.video.metric import compare
>>> from cutcutcodec.utils import get_project_root
>>> video = get_project_root() / "media" / "video" / "intro.webm"
>>> res = compare(video, video, psnr=True, ssim=True)
>>> pprint.pprint(res)  
{'psnr': [100.0,
          100.0,
          ...,
          100.0,
          100.0],
 'ssim': [1.0,
          1.0,
          ...,
          1.0,
          1.0]}
>>>

cutcutcodec.core.analysis.video.metric.lpips(ref: Tensor, dis: Tensor, *args, **kwargs) → Tensor[source]

Compute the Learned Perceptual Image Patch Similarity.

It uses the module pip install lpips in backend, based on torch.

Parameters

ref, disarraylike: The 2 images to be compared, of shape ([*batch], height, width, channels). The frames are assumed to be in RGB in range [0, 1]. Gamut and EOTF must be standard rgb.
netstr, default=”alex”: The neuronal network used, “alex” or “vgg”.
threadsint, optional: Defines the number of threads. The value -1 means that the function uses as many calculation threads as there are cores. The default value (0) allows the same behavior as (-1) if the function is called in the main thread, otherwise (1) to avoid nested threads. Any other positive value corresponds to the number of threads used.

Returns

lpipsarraylike: The learned perceptual image patch similarity of each layers.

Examples

>>> import numpy as np
>>> from cutcutcodec.core.analysis.video.metric import lpips
>>> np.random.seed(0)
>>> ref = np.random.random((720, 1080, 3))  # It could also be a torch array list...
>>> dis = 0.8 * ref + 0.2 * np.random.random((720, 1080, 3))
>>> lpips(ref, dis).round(1)
np.float64(0.0)
>>>

cutcutcodec.core.analysis.video.metric.psnr(ref: Tensor, dis: Tensor, *args, **kwargs) → Tensor[source]

Compute the peak signal to noise ratio of 2 images.

Parameters

ref, disarraylike: The 2 images to be compared, of shape ([*batch], height, width, channels). Supported types are float32 and float64.
weightsiterable[float], optional: The relative weight of each channel. By default, all channels have the same weight.
threadsint, optional: Defines the number of threads. The value -1 means that the function uses as many calculation threads as there are cores. The default value (0) allows the same behavior as (-1) if the function is called in the main thread, otherwise (1) to avoid nested threads. Any other positive value corresponds to the number of threads used.

Returns

psnrarraylike: The global peak signal to noise ratio, as a ponderation of the mean square error of each channel. It is batched and clamped in [0, 100] db.

Notes

It is optimized for C contiguous tensors.
If device is cpu and gradient is not required, a fast C code is used instead of torch code.

Examples

>>> import numpy as np
>>> from cutcutcodec.core.analysis.video.metric import psnr
>>> np.random.seed(0)
>>> ref = np.random.random((720, 1080, 3))  # It could also be a torch array list...
>>> dis = 0.8 * ref + 0.2 * np.random.random((720, 1080, 3))
>>> psnr(ref, dis).round(1)
np.float64(21.8)
>>>

cutcutcodec.core.analysis.video.metric.ssim(ref: Tensor, dis: Tensor, *args, stride: int = 1, **kwargs) → Tensor[source]

Compute the Structural similarity index measure of 2 images.

Parameters

ref, disarraylike: The 2 images to be compared, of shape ([*batch], height, width, channels). Supported types are float32 and float64.
data_rangefloat, default=1.0: The data range of the input image (difference between maximum and minimum possible values).
weightsiterable[float], optional: The relative weight of each channel. By default, all channels have the same weight.
sigmafloat, default=1.5: The standard deviation of the gaussian. It has to be strictely positive.
strideint, default=1: The stride of the convolving kernel.
threadsint, optional: Defines the number of threads. The value -1 means that the function uses as many calculation threads as there are cores. The default value (0) allows the same behavior as (-1) if the function is called in the main thread, otherwise (1) to avoid nested threads. Any other positive value corresponds to the number of threads used.

Returns

ssimarraylike: The ponderated structural similarity index measure of each layers.

Notes

It is optimized for C contiguous tensors.
If device is cpu, gradient is not required and stride != 1, a fast C code is used.

Examples

>>> import numpy as np
>>> from cutcutcodec.core.analysis.video.metric import ssim
>>> np.random.seed(0)
>>> ref = np.random.random((720, 1080, 3))  # It could also be a torch array list...
>>> dis = 0.8 * ref + 0.2 * np.random.random((720, 1080, 3))
>>> ssim(ref, dis).round(2)
np.float64(0.95)
>>>

cutcutcodec.core.analysis.video.metric.uvq(dis: Tensor, _model=None) → Tensor[source]

Compute the Perceptual Video Quality.

Parameters

disarraylike: The frames to be evaluated, of shape ([*batch], fps=5, height, width, channels=3). The framerate is assumed to be 5 Hz. The frames are assumed to be in RGB in range [0, 1]. Gamut and EOTF must be standard rgb.

Returns

uvqarraylike: The perceptual video quality measure for each group of 5 images.

Examples

>>> import numpy as np
>>> from cutcutcodec.core.analysis.video.metric import uvq
>>> np.random.seed(0)
>>> dis = np.random.random((5, 720, 1080, 3))  # It could also be a torch array list...
>>> uvq(dis).round(1)
np.float32(3.3)
>>>

Modules

`lpips_torch`	Compute a differenciable batched torch lpips.
`metric`	This module, implemented in C, offers functions for image metric calculation.
`psnr_torch`	Compute a differenciable batched torch psnr.
`ssim_torch`	Compute a differenciable batched torch ssim.
`utils`	Helper for metrics.
`uvq_google`	Universal Video Quality Model.
`vmaf`(ref, dis[, threads])	Call the Netflix vmaf metric on the frames.