File size: 9,794 Bytes
4a582ec
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# Grad and Conn is refer to https://github.com/yucornetto/MGMatting/blob/main/code-base/utils/evaluate.py
# Output of `Grad` is sightly different from the MATLAB version provided by Adobe (less than 0.1%)
# Output of `Conn` is smaller than the MATLAB version (~5%, maybe MATLAB has a different algorithm)
# So do not report results calculated by these functions in your paper.
# Evaluate your inference with the MATLAB file `DIM_evaluation_code/evaluate.m`.

import cv2
import numpy as np
from scipy.ndimage import convolve
from scipy.special import gamma
from skimage.measure import label


class MSE:
    """
    Only calculate the unknown region if trimap provided.
    """

    def __init__(self):
        self.mse_diffs = 0
        self.count = 0

    def update(self, pred, gt, trimap=None):
        """
        update metric.
        Args:
            pred (np.ndarray): The value range is [0., 255.].
            gt (np.ndarray): The value range is [0, 255].
            trimap (np.ndarray, optional) The value is in {0, 128, 255}. Default: None.
        """
        if trimap is None:
            trimap = np.ones_like(gt) * 128
        if not (pred.shape == gt.shape == trimap.shape):
            raise ValueError(
                'The shape of `pred`, `gt` and `trimap` should be equal. '
                'but they are {}, {} and {}'.format(pred.shape, gt.shape,
                                                    trimap.shape))
        pred[trimap == 0] = 0
        pred[trimap == 255] = 255

        mask = trimap == 128
        pixels = float(mask.sum())
        pred = pred / 255.
        gt = gt / 255.
        diff = (pred - gt) * mask
        mse_diff = (diff**2).sum() / pixels if pixels > 0 else 0

        self.mse_diffs += mse_diff
        self.count += 1

        return mse_diff

    def evaluate(self):
        mse = self.mse_diffs / self.count if self.count > 0 else 0
        return mse


class SAD:
    """
    Only calculate the unknown region if trimap provided.
    """

    def __init__(self):
        self.sad_diffs = 0
        self.count = 0

    def update(self, pred, gt, trimap=None):
        """
        update metric.
        Args:
            pred (np.ndarray): The value range is [0., 255.].
            gt (np.ndarray): The value range is [0., 255.].
            trimap (np.ndarray, optional)L The value is in {0, 128, 255}. Default: None.
        """
        if trimap is None:
            trimap = np.ones_like(gt) * 128
        if not (pred.shape == gt.shape == trimap.shape):
            raise ValueError(
                'The shape of `pred`, `gt` and `trimap` should be equal. '
                'but they are {}, {} and {}'.format(pred.shape, gt.shape,
                                                    trimap.shape))
        pred[trimap == 0] = 0
        pred[trimap == 255] = 255

        mask = trimap == 128
        pred = pred / 255.
        gt = gt / 255.
        diff = (pred - gt) * mask
        sad_diff = (np.abs(diff)).sum()

        sad_diff /= 1000
        self.sad_diffs += sad_diff
        self.count += 1

        return sad_diff

    def evaluate(self):
        sad = self.sad_diffs / self.count if self.count > 0 else 0
        return sad


class Grad:
    """
    Only calculate the unknown region if trimap provided.
    Refer to: https://github.com/open-mlab/mmediting/blob/master/mmedit/core/evaluation/metrics.py
    """

    def __init__(self):
        self.grad_diffs = 0
        self.count = 0

    def gaussian(self, x, sigma):
        return np.exp(-x**2 / (2 * sigma**2)) / (sigma * np.sqrt(2 * np.pi))

    def dgaussian(self, x, sigma):
        return -x * self.gaussian(x, sigma) / sigma**2

    def gauss_filter(self, sigma, epsilon=1e-2):
        half_size = np.ceil(
            sigma * np.sqrt(-2 * np.log(np.sqrt(2 * np.pi) * sigma * epsilon)))
        size = int(2 * half_size + 1)

        # create filter in x axis
        filter_x = np.zeros((size, size))
        for i in range(size):
            for j in range(size):
                filter_x[i, j] = self.gaussian(
                    i - half_size, sigma) * self.dgaussian(j - half_size, sigma)

        # normalize filter
        norm = np.sqrt((filter_x**2).sum())
        filter_x = filter_x / norm
        filter_y = np.transpose(filter_x)

        return filter_x, filter_y

    def gauss_gradient(self, img, sigma):
        filter_x, filter_y = self.gauss_filter(sigma)
        img_filtered_x = cv2.filter2D(
            img, -1, filter_x, borderType=cv2.BORDER_REPLICATE)
        img_filtered_y = cv2.filter2D(
            img, -1, filter_y, borderType=cv2.BORDER_REPLICATE)
        return np.sqrt(img_filtered_x**2 + img_filtered_y**2)

    def update(self, pred, gt, trimap=None, sigma=1.4):
        """
        update metric.
        Args:
            pred (np.ndarray): The value range is [0., 1.].
            gt (np.ndarray): The value range is [0, 255].
            trimap (np.ndarray, optional)L The value is in {0, 128, 255}. Default: None.
            sigma (float, optional): Standard deviation of the gaussian kernel. Default: 1.4.
        """
        if trimap is None:
            trimap = np.ones_like(gt) * 128
        if not (pred.shape == gt.shape == trimap.shape):
            raise ValueError(
                'The shape of `pred`, `gt` and `trimap` should be equal. '
                'but they are {}, {} and {}'.format(pred.shape, gt.shape,
                                                    trimap.shape))
        pred[trimap == 0] = 0
        pred[trimap == 255] = 255

        gt = gt.squeeze()
        pred = pred.squeeze()
        gt = gt.astype(np.float64)
        pred = pred.astype(np.float64)
        gt_normed = np.zeros_like(gt)
        pred_normed = np.zeros_like(pred)
        cv2.normalize(gt, gt_normed, 1., 0., cv2.NORM_MINMAX)
        cv2.normalize(pred, pred_normed, 1., 0., cv2.NORM_MINMAX)

        gt_grad = self.gauss_gradient(gt_normed, sigma).astype(np.float32)
        pred_grad = self.gauss_gradient(pred_normed, sigma).astype(np.float32)

        grad_diff = ((gt_grad - pred_grad)**2 * (trimap == 128)).sum()

        grad_diff /= 1000
        self.grad_diffs += grad_diff
        self.count += 1

        return grad_diff

    def evaluate(self):
        grad = self.grad_diffs / self.count if self.count > 0 else 0
        return grad


class Conn:
    """
    Only calculate the unknown region if trimap provided.
    Refer to: Refer to: https://github.com/open-mlab/mmediting/blob/master/mmedit/core/evaluation/metrics.py
    """

    def __init__(self):
        self.conn_diffs = 0
        self.count = 0

    def update(self, pred, gt, trimap=None, step=0.1):
        """
        update metric.
        Args:
            pred (np.ndarray): The value range is [0., 1.].
            gt (np.ndarray): The value range is [0, 255].
            trimap (np.ndarray, optional)L The value is in {0, 128, 255}. Default: None.
            step (float, optional): Step of threshold when computing intersection between
            `gt` and `pred`. Default: 0.1.
        """
        if trimap is None:
            trimap = np.ones_like(gt) * 128
        if not (pred.shape == gt.shape == trimap.shape):
            raise ValueError(
                'The shape of `pred`, `gt` and `trimap` should be equal. '
                'but they are {}, {} and {}'.format(pred.shape, gt.shape,
                                                    trimap.shape))
        pred[trimap == 0] = 0
        pred[trimap == 255] = 255

        gt = gt.squeeze()
        pred = pred.squeeze()
        gt = gt.astype(np.float32) / 255
        pred = pred.astype(np.float32) / 255

        thresh_steps = np.arange(0, 1 + step, step)
        round_down_map = -np.ones_like(gt)
        for i in range(1, len(thresh_steps)):
            gt_thresh = gt >= thresh_steps[i]
            pred_thresh = pred >= thresh_steps[i]
            intersection = (gt_thresh & pred_thresh).astype(np.uint8)

            # connected components
            _, output, stats, _ = cv2.connectedComponentsWithStats(
                intersection, connectivity=4)
            # start from 1 in dim 0 to exclude background
            size = stats[1:, -1]

            # largest connected component of the intersection
            omega = np.zeros_like(gt)
            if len(size) != 0:
                max_id = np.argmax(size)
                # plus one to include background
                omega[output == max_id + 1] = 1

            mask = (round_down_map == -1) & (omega == 0)
            round_down_map[mask] = thresh_steps[i - 1]
        round_down_map[round_down_map == -1] = 1

        gt_diff = gt - round_down_map
        pred_diff = pred - round_down_map
        # only calculate difference larger than or equal to 0.15
        gt_phi = 1 - gt_diff * (gt_diff >= 0.15)
        pred_phi = 1 - pred_diff * (pred_diff >= 0.15)

        conn_diff = np.sum(np.abs(gt_phi - pred_phi) * (trimap == 128))

        conn_diff /= 1000
        self.conn_diffs += conn_diff
        self.count += 1

        return conn_diff

    def evaluate(self):
        conn = self.conn_diffs / self.count if self.count > 0 else 0
        return conn