nerfbaselines.metrics

nerfbaselines.metrics.dmpix_ssim(a: ndarray, b: ndarray, *, max_val: float = 1.0, kernel_size: int = 11, sigma: float = 1.5, k1: float = 0.01, k2: float = 0.03, return_map: bool = False, filter_fn: Callable[[ndarray], ndarray] | None = None) → ndarray[source]

Computes the structural similarity index (SSIM) between image pairs.

This function is based on the standard SSIM implementation from: Z. Wang, A. C. Bovik, H. R. Sheikh and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity”, in IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600-612, 2004.

This function was modeled after tf.image.ssim, and should produce comparable output.

Note: the true SSIM is only defined on grayscale. This function does not perform any colorspace transform. If the input is in a color space, then it will compute the average SSIM.

Note

This function exactly matches dm_pix.ssim

Parameters:

  • a – First image (or set of images).

  • b – Second image (or set of images).

  • max_val – The maximum magnitude that a or b can have.

  • kernel_size – Window size (>= 1). Image dims must be at least this small.

  • sigma – The bandwidth of the Gaussian used for filtering (> 0.).

  • k1 – One of the SSIM dampening parameters (> 0.).

  • k2 – One of the SSIM dampening parameters (> 0.).

  • return_map – If True, will cause the per-pixel SSIM “map” to be returned.

  • precision – The numerical precision to use when performing convolution.

Returns:

Each image’s mean SSIM, or a tensor of individual values if return_map.

nerfbaselines.metrics.lpips(a: ndarray, b: ndarray) → ndarray | float32[source]

Compute Learned Perceptual Image Patch Similarity (the lower the better). :param a: Tensor of prediction images [B…, H, W, C]. :param b: Tensor of target images [B…, H, W, C].

Returns:

Tensor of LPIPS values for each image [B…].

nerfbaselines.metrics.lpips_alex(a: ndarray, b: ndarray) → ndarray | float32[source]

Compute Learned Perceptual Image Patch Similarity (the lower the better). :param a: Tensor of prediction images [B…, H, W, C]. :param b: Tensor of target images [B…, H, W, C].

Returns:

Tensor of LPIPS values for each image [B…].

nerfbaselines.metrics.lpips_vgg(a: ndarray, b: ndarray) → ndarray | float32[source]

Compute Learned Perceptual Image Patch Similarity (the lower the better). :param a: Tensor of prediction images [B…, H, W, C]. :param b: Tensor of target images [B…, H, W, C].

Returns:

Tensor of LPIPS values for each image [B…].

nerfbaselines.metrics.mae(a: ndarray, b: ndarray) → ndarray | float32[source]

Compute Mean Absolute Error (the lower the better). :param a: Tensor of prediction images [B, H, W, C]. :param b: Tensor of target images [B, H, W, C].

Returns:

Tensor of mean absolute error values for each image [B].

nerfbaselines.metrics.mse(a: ndarray, b: ndarray) → ndarray | float32[source]

Compute Mean Squared Error (the lower the better). :param a: Tensor of prediction images [B, H, W, C]. :param b: Tensor of target images [B, H, W, C].

Returns:

Tensor of mean squared error values for each image [B].

nerfbaselines.metrics.psnr(a: ndarray | float32 | float64, b: ndarray | None = None) → ndarray | float32 | float64[source]

Compute Peak Signal to Noise Ratio (the higher the better). It can reuse computed MSE values if b is None. :param a: Tensor of prediction images [B, H, W, C] or a tensor of MSE values [B] (b must be None in that case). :param b: Tensor of target images [B, H, W, C] or None (if a are MSE values).

Returns:

Tensor of PSNR values for each image [B].

nerfbaselines.metrics.ssim(a: ndarray, b: ndarray) → ndarray | float32[source]

Compute Structural Similarity Index Measure (the higher the better). :param a: Tensor of prediction images [B, H, W, C]. :param b: Tensor of target images [B, H, W, C].

Returns:

Tensor of mean SSIM values for each image [B].

nerfbaselines.metrics.torchmetrics_ssim(a: ndarray, b: ndarray, *, gaussian_kernel: bool = True, sigma: float | Sequence[float] = 1.5, kernel_size: int | Sequence[int] = 11, data_range: float | Tuple[float, float] | None = None, k1: float = 0.01, k2: float = 0.03) → ndarray | Tuple[ndarray, ndarray][source]

Compute Structural Similarity Index Measure.

Note

This metric exactly matches torchmetrics.ssim

Parameters:

  • preds – estimated image

  • target – ground truth image

  • gaussian_kernel – If true (default), a gaussian kernel is used, if false a uniform kernel is used

  • sigma – Standard deviation of the gaussian kernel, anisotropic kernels are possible. Ignored if a uniform kernel is used

  • kernel_size – the size of the uniform kernel, anisotropic kernels are possible. Ignored if a Gaussian kernel is used

  • data_range – Range of the image. If None, it is determined from the image (max - min)

  • k1 – Parameter of SSIM.

  • k2 – Parameter of SSIM.