File size: 12,026 Bytes
b9e72b4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
import math
import os.path
import os.path
import random
from os.path import join

import cv2
import numpy as np
import torch.utils.data
import torchvision.transforms.functional as TF #more flexible and fine-grained transformations compared to the standard transforms API.
from PIL import Image
from scipy.signal import convolve2d

from data.image_folder import make_dataset
from data.torchdata import Dataset as BaseDataset #overrides len(), + and []
from data.transforms import to_tensor


def __scale_width(img, target_width):
    ow, oh = img.size #img.size method from (PIL) /Pillow, returns a tuple (width,height) in px
    #original width and original height
    if (ow == target_width):
        return img
    w = target_width
    h = int(target_width * oh / ow) #maintain aspect ratio by changing height
    h = math.ceil(h / 2.) * 2  #(mingcv) round up to even
    return img.resize((w, h), Image.BICUBIC) #resizes the image to the new dimensions (w, h) using bicubic interpolation. Bicubic interpolation is a method of resizing images that results in smoother and more visually appealing images compared to simpler methods like nearest neighbor or bilinear interpolation. From PIL/pillow


def __scale_height(img, target_height):
    ow, oh = img.size
    if (oh == target_height):
        return img
    h = target_height
    w = int(target_height * ow / oh)
    w = math.ceil(w / 2.) * 2
    return img.resize((w, h), Image.BICUBIC)


def paired_data_transforms(img_1, img_2, unaligned_transforms=False):#img1, img2:PIL Image objects 
#Unaligned transformations refer to transformations applied to paired images where the transformations are intentionally not synchronized between the two images. This approach introduces a controlled misalignment, adding variability to the data and potentially increasing the robustness of a model trained on such data.
    def get_params(img, output_size):#inner nested function for scope & encapsulation
        #getting dimensions for cropping
        w, h = img.size
        th, tw = output_size
        if w == tw and h == th:
            return 0, 0, h, w# no cropping as u crop form top left to bottom right

        i = random.randint(0, h - th)
        j = random.randint(0, w - tw)
        return i, j, th, tw #i,j = starting position of cropped region

    target_size = int(random.randint(224, 448) / 2.) * 2 
    '''??? why 224-448'''
    ow, oh = img_1.size
    if ow >= oh:#scale the smaller dimension
        img_1 = __scale_height(img_1, target_size)
        img_2 = __scale_height(img_2, target_size)
    else:
        img_1 = __scale_width(img_1, target_size)
        img_2 = __scale_width(img_2, target_size)

    if random.random() < 0.5: #flip randomly pairs
        img_1 = TF.hflip(img_1)
        img_2 = TF.hflip(img_2)

    if random.random() < 0.5: #rotate randomly pairs
        angle = random.choice([90, 180, 270])
        img_1 = TF.rotate(img_1, angle)
        img_2 = TF.rotate(img_2, angle)
    #unalined transoformation by cropping 2 images differently
    i, j, h, w = get_params(img_1, (224, 224))
    img_1 = TF.crop(img_1, i, j, h, w) 

    if unaligned_transforms:
        #(mingcv) print('random shift')
        i_shift = random.randint(-10, 10)
        j_shift = random.randint(-10, 10)
        i += i_shift
        j += j_shift

    img_2 = TF.crop(img_2, i, j, h, w)

    return img_1, img_2


class ReflectionSynthesis(object):
    def __init__(self):
        #(mingcv) Kernel Size of the Gaussian Blurry
        self.kernel_sizes = [5, 7, 9, 11]
        self.kernel_probs = [0.1, 0.2, 0.3, 0.4]

        #(mingcv) Sigma of the Gaussian Blurry
        self.sigma_range = [2, 5]
        self.alpha_range = [0.8, 1.0]
        self.beta_range = [0.4, 1.0]

    def __call__(self, T_, R_):
        T_ = np.asarray(T_, np.float32) / 255.
        R_ = np.asarray(R_, np.float32) / 255.

        kernel_size = np.random.choice(self.kernel_sizes, p=self.kernel_probs)
        sigma = np.random.uniform(self.sigma_range[0], self.sigma_range[1])
        kernel = cv2.getGaussianKernel(kernel_size, sigma)
        kernel2d = np.dot(kernel, kernel.T)
        for i in range(3):
            R_[..., i] = convolve2d(R_[..., i], kernel2d, mode='same')

        a = np.random.uniform(self.alpha_range[0], self.alpha_range[1])
        b = np.random.uniform(self.beta_range[0], self.beta_range[1])
        T, R = a * T_, b * R_

        if random.random() < 0.7:
            I = T + R - T * R
        else:
            I = T + R
            if np.max(I) > 1:
                m = I[I > 1]
                m = (np.mean(m) - 1) * 1.3
                I = np.clip(T + np.clip(R - m, 0, 1), 0, 1)

        return T_, R_, I


class DataLoader(torch.utils.data.DataLoader):
    def __init__(self, dataset, batch_size, shuffle, *args, **kwargs):
        super(DataLoader, self).__init__(dataset, batch_size, shuffle, *args, **kwargs)
        self.shuffle = shuffle

    def reset(self):
        if self.shuffle:
            print('Reset Dataset...')
            self.dataset.reset()


class DSRDataset(BaseDataset):
    def __init__(self, datadir, fns=None, size=None, enable_transforms=True):
        super(DSRDataset, self).__init__()
        self.size = size
        self.datadir = datadir
        self.enable_transforms = enable_transforms
        sortkey = lambda key: os.path.split(key)[-1]
        self.paths = sorted(make_dataset(datadir, fns), key=sortkey)
        if size is not None:
            self.paths = np.random.choice(self.paths, size)

        self.syn_model = ReflectionSynthesis()
        self.reset(shuffle=False)

    def reset(self, shuffle=True):
        if shuffle:
            random.shuffle(self.paths)
        num_paths = len(self.paths) // 2
        self.B_paths = self.paths[0:num_paths]
        self.R_paths = self.paths[num_paths:2 * num_paths]

    def data_synthesis(self, t_img, r_img):
        if self.enable_transforms:
            t_img, r_img = paired_data_transforms(t_img, r_img)

        t_img, r_img, m_img = self.syn_model(t_img, r_img)

        B = TF.to_tensor(t_img)
        R = TF.to_tensor(r_img)
        M = TF.to_tensor(m_img)

        return B, R, M

    def __getitem__(self, index):
        index_B = index % len(self.B_paths)
        index_R = index % len(self.R_paths)

        B_path = self.B_paths[index_B]
        R_path = self.R_paths[index_R]

        t_img = Image.open(B_path).convert('RGB')
        r_img = Image.open(R_path).convert('RGB')

        B, R, M = self.data_synthesis(t_img, r_img)
        fn = os.path.basename(B_path)
        return {'input': M, 'target_t': B, 'target_r': R, 'fn': fn, 'real': False}

    def __len__(self):
        if self.size is not None:
            return min(max(len(self.B_paths), len(self.R_paths)), self.size)
        else:
            return max(len(self.B_paths), len(self.R_paths))


class DSRTestDataset(BaseDataset):
    def __init__(self, datadir, fns=None, size=None, enable_transforms=False, unaligned_transforms=False,

                 round_factor=1, flag=None, if_align=False):
        super(DSRTestDataset, self).__init__()
        self.size = size
        self.datadir = datadir
        self.fns = fns or os.listdir(join(datadir, 'blended'))
        self.enable_transforms = enable_transforms
        self.unaligned_transforms = unaligned_transforms
        self.round_factor = round_factor
        self.flag = flag
        self.if_align = if_align

        if size is not None:
            self.fns = self.fns[:size]

    def align(self, x1, x2):
        h, w = x1.height, x1.width
        h, w = h // 32 * 32, w // 32 * 32
        x1 = x1.resize((w, h))
        x2 = x2.resize((w, h))
        return x1, x2

    def __getitem__(self, index):
        fn = self.fns[index]

        t_img = Image.open(join(self.datadir, 'transmission_layer', fn)).convert('RGB')
        m_img = Image.open(join(self.datadir, 'blended', fn)).convert('RGB')

        if self.if_align:
            t_img, m_img = self.align(t_img, m_img)

        if self.enable_transforms:
            t_img, m_img = paired_data_transforms(t_img, m_img, self.unaligned_transforms)

        B = TF.to_tensor(t_img)
        M = TF.to_tensor(m_img)

        dic = {'input': M, 'target_t': B, 'fn': fn, 'real': True, 'target_r': M - B}
        if self.flag is not None:
            dic.update(self.flag)
        return dic

    def __len__(self):
        if self.size is not None:
            return min(len(self.fns), self.size)
        else:
            return len(self.fns)


class SIRTestDataset(BaseDataset):
    def __init__(self, datadir, fns=None, size=None, if_align=False):
        super(SIRTestDataset, self).__init__()
        self.size = size
        self.datadir = datadir
        self.fns = fns or os.listdir(join(datadir, 'blended'))
        self.if_align = if_align

        if size is not None:
            self.fns = self.fns[:size]

    def align(self, x1, x2, x3):
        h, w = x1.height, x1.width
        h, w = h // 32 * 32, w // 32 * 32
        x1 = x1.resize((w, h))
        x2 = x2.resize((w, h))
        x3 = x3.resize((w, h))
        return x1, x2, x3

    def __getitem__(self, index):
        fn = self.fns[index]

        t_img = Image.open(join(self.datadir, 'transmission_layer', fn)).convert('RGB')
        r_img = Image.open(join(self.datadir, 'reflection_layer', fn)).convert('RGB')
        m_img = Image.open(join(self.datadir, 'blended', fn)).convert('RGB')

        if self.if_align:
            t_img, r_img, m_img = self.align(t_img, r_img, m_img)

        B = TF.to_tensor(t_img)
        R = TF.to_tensor(r_img)
        M = TF.to_tensor(m_img)

        dic = {'input': M, 'target_t': B, 'fn': fn, 'real': True, 'target_r': R, 'target_r_hat': M - B}
        return dic

    def __len__(self):
        if self.size is not None:
            return min(len(self.fns), self.size)
        else:
            return len(self.fns)


class RealDataset(BaseDataset):
    def __init__(self, datadir, fns=None, size=None):
        super(RealDataset, self).__init__()
        self.size = size
        self.datadir = datadir
        self.fns = fns or os.listdir(join(datadir))

        if size is not None:
            self.fns = self.fns[:size]

    def align(self, x):
        h, w = x.height, x.width
        h, w = h // 32 * 32, w // 32 * 32
        x = x.resize((w, h))
        return x

    def __getitem__(self, index):
        fn = self.fns[index]
        B = -1
        m_img = Image.open(join(self.datadir, fn)).convert('RGB')
        M = to_tensor(self.align(m_img))
        data = {'input': M, 'target_t': B, 'fn': fn}
        return data

    def __len__(self):
        if self.size is not None:
            return min(len(self.fns), self.size)
        else:
            return len(self.fns)


class FusionDataset(BaseDataset):
    def __init__(self, datasets, fusion_ratios=None, size=None):
        self.datasets = datasets
        self.size = size or sum([len(dataset) for dataset in datasets])
        self.fusion_ratios = fusion_ratios or [1. / len(datasets)] * len(datasets)
        print('[i] using a fusion dataset: %d %s imgs fused with ratio %s' % (
            self.size, [len(dataset) for dataset in datasets], self.fusion_ratios))

    def reset(self):
        for dataset in self.datasets:
            dataset.reset()

    def __getitem__(self, index):
        residual = 1
        for i, ratio in enumerate(self.fusion_ratios):
            if random.random() < ratio / residual or i == len(self.fusion_ratios) - 1:
                dataset = self.datasets[i]
                return dataset[index % len(dataset)]
            residual -= ratio

    def __len__(self):
        return self.size