[Feature] Support SWD and MS-SSIM metrics

mmeval/metrics/__init__.py

@@ -20,6 +20,7 @@
 from .mae import MeanAbsoluteError
 from .matting_mse import MattingMeanSquaredError
 from .mean_iou import MeanIoU
+from .ms_ssim import MultiScaleStructureSimilarity
 from .mse import MeanSquaredError
 from .niqe import NaturalImageQualityEvaluator
 from .oid_map import OIDMeanAP
@@ -34,6 +35,7 @@
 from .sad import SumAbsoluteDifferences
 from .snr import SignalNoiseRatio
 from .ssim import StructuralSimilarity
+from .swd import SlicedWassersteinDistance
 from .voc_map import VOCMeanAP
 from .word_accuracy import WordAccuracy
 
@@ -48,7 +50,8 @@
     'KeypointAUC', 'KeypointNME', 'NaturalImageQualityEvaluator',
     'WordAccuracy', 'PrecisionRecallF1score',
     'SingleLabelPrecisionRecallF1score', 'MultiLabelPrecisionRecallF1score',
-    'CharRecallPrecision'
+    'CharRecallPrecision', 'MultiScaleStructureSimilarity',
+    'SlicedWassersteinDistance'
 ]
 
 _deprecated_msg = (

mmeval/metrics/ms_ssim.py

@@ -0,0 +1,295 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import numpy as np
+from scipy import signal
+from typing import Dict, List, Sequence, Tuple
+
+from mmeval.core import BaseMetric
+from .utils.image_transforms import reorder_image
+
+
+class MultiScaleStructureSimilarity(BaseMetric):
+    """MS-SSIM (Multi-Scale Structure Similarity) metric.
+
+    Ref:
+    This class implements Multi-Scale Structural Similarity (MS-SSIM) Image
+    Quality Assessment according to Zhou Wang's paper, "Multi-scale structural
+    similarity for image quality assessment" (2003).
+    Link: https://ece.uwaterloo.ca/~z70wang/publications/msssim.pdf
+
+    Author's MATLAB implementation:
+    http://www.cns.nyu.edu/~lcv/ssim/msssim.zip
+
+    PGGAN's implementation:
+    https://github.com/tkarras/progressive_growing_of_gans/blob/master/metrics/ms_ssim.py
+
+    Args:
+        input_order (str): Whether the input order is 'HWC' or 'CHW'.
+            Defaults to 'HWC'.
+        max_val (int): the dynamic range of the images (i.e., the difference
+            between the maximum and the minimum allowed values).
+            Defaults to 255.
+        filter_size (int): Size of blur kernel to use (will be reduced for
+            small images). Defaults to 11.
+        filter_sigma (float): Standard deviation for Gaussian blur kernel (will
+            be reduced for small images). Defaults to 1.5.
+        k1 (float): Constant used to maintain stability in the SSIM calculation
+            (0.01 in the original paper). Defaults to 0.01.
+        k2 (float): Constant used to maintain stability in the SSIM calculation
+            (0.03 in the original paper). Defaults to 0.03.
+        weights (List[float]): List of weights for each level. Defaults to
+            [0.0448, 0.2856, 0.3001, 0.2363, 0.1333]. Noted that the default
+            weights don't sum to 1.0 but do match the paper / matlab code.
+        **kwargs: Keyword parameters passed to :class:`BaseMetric`.
+
+    Examples:
+
+        >>> from mmeval import MultiScaleStructureSimilarity as MS_SSIM
+        >>> import numpy as np
+        >>>
+        >>> ms_ssim = MS_SSIM()
+        >>> preds = [np.random.randint(0, 255, size=(3, 32, 32)) for _ in range(4)]  # noqa
+        >>> gts = [np.random.randint(0, 255, size=(3, 32, 32)) for _ in range(4)]  # noqa
+        >>> ms_ssim(preds, gts)  # doctest: +ELLIPSIS
+        {'ms_ssim': ...}
+    """
+
+    def __init__(self,
+                 input_order: str = 'CHW',
+                 max_val: int = 255,
+                 filter_size: int = 11,
+                 filter_sigma: float = 1.5,
+                 k1: float = 0.01,
+                 k2: float = 0.03,
+                 weights: List[float] = [
+                     0.0448, 0.2856, 0.3001, 0.2363, 0.1333
+                 ],
+                 **kwargs) -> None:
+        super().__init__(**kwargs)
+
+        assert input_order.upper() in [
+            'CHW', 'HWC'
+        ], (f'Wrong input_order {input_order}. Supported input_orders are '
+            '"HWC" and "CHW"')
+        self.input_order = input_order
+
+        self.max_val = max_val
+        self.filter_size = filter_size
+        self.filter_sigma = filter_sigma
+        self.k1 = k1
+        self.k2 = k2
+        self.weights = np.array(weights)
+
+    def add(self, predictions: Sequence[np.ndarray], groundtruths: Sequence[np.ndarray]) -> None:  # type: ignore # yapf: disable # noqa: E501
+        """Add a bunch of images to calculate metric result.
+
+        Args:
+            predictions (Sequence[np.ndarray]): Predictions of the model.
+                The length of `predictions` must be same as `groundtruths`.
+                The width and height of each element must be divisible by 2 **
+                num_scale (`self.weights.size`). The channel order of each
+                element should align with `self.input_order` and the range
+                should be [0, 255].
+            groundtruths (Sequence[np.ndarray], optional): Groundtruth of the
+                model. The number of elements in the Sequence must be same as
+                `predictions`, and the width and height of each element must
+                be divisible by 2 ** num_scale (`self.weights.size`). The
+                channel order of each element should align with
+                `self.input_order` and the range should be [0, 255].
+                Defaults to None.
+        """
+        assert len(predictions) == len(groundtruths), (
+            'The length of "predictions" and "groundtruths" must be '
+            'same.')
+        half1, half2 = predictions, groundtruths
+
+        half1 = [reorder_image(samp, self.input_order) for samp in half1]
+        half2 = [reorder_image(samp, self.input_order) for samp in half2]
+        least_size = 2**self.weights.size
+        assert all([
+            sample.shape[0] % least_size == 0 for sample in half1
+        ]), ('The height and width of each sample must be divisible by '
+             f'{least_size} (2 ** len(self.weights.size)).')
+        assert all([
+            sample.shape[0] % least_size == 0 for sample in half2
+        ]), ('The height and width of each sample must be divisible by '
+             f'{least_size} (2 ** self.weights.size).')
+
+        half1 = np.stack(half1, axis=0).astype(np.uint8)
+        half2 = np.stack(half2, axis=0).astype(np.uint8)
+
+        self._results += self.compute_ms_ssim(half1, half2)
+
+    def compute_metric(self, results: List[np.float64]) -> Dict[str, float]:
+        """Compute the MS-SSIM metric.
+
+        This method would be invoked in ``BaseMetric.compute`` after
+        distributed synchronization.
+
+        Args:
+            results (List[np.float64]): A list that consisting the PSNR score.
+                This list has already been synced across all ranks.
+
+        Returns:
+            Dict[str, float]: The computed PSNR metric.
+        """
+        return {'ms-ssim': float(np.array(results).mean())}
+
+    def compute_ms_ssim(self, img1: np.array, img2: np.array) -> List[float]:
+        """Calculate MS-SSIM (multi-scale structural similarity).
+
+        Args:
+            img1 (ndarray): Images with range [0, 255] and order "NHWC".
+            img2 (ndarray): Images with range [0, 255] and order "NHWC".
+
+        Returns:
+            np.ndarray: MS-SSIM score between `img1` and `img2` of shape (N, ).
+        """
+        if img1.shape != img2.shape:
+            raise RuntimeError(
+                'Input images must have the same shape (%s vs. %s).' %
+                (img1.shape, img2.shape))
+        if img1.ndim != 4:
+            raise RuntimeError(
+                'Input images must have four dimensions, not %d' % img1.ndim)
+
+        levels = self.weights.size
+        im1, im2 = (x.astype(np.float32) for x in [img1, img2])
+        mssim = []
+        mcs = []
+        for _ in range(levels):
+            ssim, cs = self._ssim_for_multi_scale(
+                im1,
+                im2,
+                max_val=self.max_val,
+                filter_size=self.filter_size,
+                filter_sigma=self.filter_sigma,
+                k1=self.k1,
+                k2=self.k2)
+            mssim.append(ssim)
+            mcs.append(cs)
+            im1, im2 = (self._hox_downsample(x) for x in [im1, im2])
+
+        # Clip to zero. Otherwise we get NaNs.
+        mssim = np.clip(np.asarray(mssim), 0.0, np.inf)
+        mcs = np.clip(np.asarray(mcs), 0.0, np.inf)
+
+        results = np.prod(
+            mcs[:-1, :]**self.weights[:-1, np.newaxis], axis=0) * (
+                mssim[-1, :]**self.weights[-1])
+        return results.tolist()
+
+    @staticmethod
+    def _f_special_gauss(size: int, sigma: float) -> np.ndarray:
+        r"""Return a circular symmetric gaussian kernel.
+
+        Ref: https://github.com/tkarras/progressive_growing_of_gans/blob/2504c3f3cb98ca58751610ad61fa1097313152bd/metrics/ms_ssim.py#L25-L36  # noqa
+
+        Args:
+            size (int): Size of Gaussian kernel.
+            sigma (float): Standard deviation for Gaussian blur kernel.
+
+        Returns:
+            np.ndarray: Gaussian kernel.
+        """
+        radius = size // 2
+        offset = 0.0
+        start, stop = -radius, radius + 1
+        if size % 2 == 0:
+            offset = 0.5
+            stop -= 1
+        x, y = np.mgrid[offset + start:stop,  # type: ignore # noqa
+                        offset + start:stop]  # type: ignore # noqa
+        assert len(x) == size
+        g = np.exp(-((x**2 + y**2) / (2.0 * sigma**2)))
+        return g / g.sum()
+
+    @staticmethod
+    def _hox_downsample(img: np.ndarray) -> np.ndarray:
+        r"""Downsample images with factor equal to 0.5.
+
+        Ref: https://github.com/tkarras/progressive_growing_of_gans/blob/2504c3f3cb98ca58751610ad61fa1097313152bd/metrics/ms_ssim.py#L110-L111  # noqa
+
+        Args:
+            img (np.ndarray): Images with order "NHWC".
+
+        Returns:
+            np.ndarray: Downsampled images with order "NHWC".
+        """
+        return (img[:, 0::2, 0::2, :] + img[:, 1::2, 0::2, :] +
+                img[:, 0::2, 1::2, :] + img[:, 1::2, 1::2, :]) * 0.25
+
+    def _ssim_for_multi_scale(
+            self,
+            img1: np.ndarray,
+            img2: np.ndarray,
+            max_val: int = 255,
+            filter_size: int = 11,
+            filter_sigma: float = 1.5,
+            k1: float = 0.01,
+            k2: float = 0.03) -> Tuple[np.ndarray, np.ndarray]:
+        """Calculate SSIM (structural similarity) and contrast sensitivity.
+
+        Ref:
+        Our implementation is based on PGGAN:
+        https://github.com/tkarras/progressive_growing_of_gans/blob/2504c3f3cb98ca58751610ad61fa1097313152bd/metrics/ms_ssim.py#L38-L108  # noqa
+
+        Args:
+            img1 (np.ndarray): Images with range [0, 255] and order "NHWC".
+            img2 (np.ndarray): Images with range [0, 255] and order "NHWC".
+            max_val (int): the dynamic range of the images (i.e., the
+                difference between the maximum the and minimum allowed
+                values). Defaults to 255.
+            filter_size (int): Size of blur kernel to use (will be reduced for
+                small images). Defaults to 11.
+            filter_sigma (float): Standard deviation for Gaussian blur kernel (
+                will be reduced for small images). Defaults to 1.5.
+            k1 (float): Constant used to maintain stability in the SSIM
+                calculation (0.01 in the original paper). Defaults to 0.01.
+            k2 (float): Constant used to maintain stability in the SSIM
+                calculation (0.03 in the original paper). Defaults to 0.03.
+
+        Returns:
+            tuple: Pair containing the mean SSIM and contrast sensitivity
+                between `img1` and `img2`.
+        """
+        img1 = img1.astype(np.float32)
+        img2 = img2.astype(np.float32)
+        _, height, width, _ = img1.shape
+
+        # Filter size can't be larger than height or width of images.
+        size = min(filter_size, height, width)
+
+        # Scale down sigma if a smaller filter size is used.
+        sigma = size * filter_sigma / filter_size if filter_size else 0
+
+        if filter_size:
+            window = np.reshape(
+                self._f_special_gauss(size, sigma), (1, size, size, 1))
+            mu1 = signal.fftconvolve(img1, window, mode='valid')
+            mu2 = signal.fftconvolve(img2, window, mode='valid')
+            sigma11 = signal.fftconvolve(img1 * img1, window, mode='valid')
+            sigma22 = signal.fftconvolve(img2 * img2, window, mode='valid')
+            sigma12 = signal.fftconvolve(img1 * img2, window, mode='valid')
+        else:
+            # Empty blur kernel so no need to convolve.
+            mu1, mu2 = img1, img2
+            sigma11 = img1 * img1
+            sigma22 = img2 * img2
+            sigma12 = img1 * img2
+
+        mu11 = mu1 * mu1
+        mu22 = mu2 * mu2
+        mu12 = mu1 * mu2
+        sigma11 -= mu11
+        sigma22 -= mu22
+        sigma12 -= mu12
+
+        # Calculate intermediate values used by both ssim and cs_map.
+        c1 = (k1 * max_val)**2
+        c2 = (k2 * max_val)**2
+        v1 = 2.0 * sigma12 + c2
+        v2 = sigma11 + sigma22 + c2
+        ssim = np.mean((((2.0 * mu12 + c1) * v1) / ((mu11 + mu22 + c1) * v2)),
+                       axis=(1, 2, 3))  # Return for each image individually.
+        cs = np.mean(v1 / v2, axis=(1, 2, 3))
+        return ssim, cs