File size: 12,029 Bytes
a80d6bb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
337
338
339
340
341
342
343
# From https://github.com/TRI-ML/KP2D.

# Copyright 2020 Toyota Research Institute.  All rights reserved.

import random
from math import pi

import cv2
import numpy as np
import torch
import torchvision
import torchvision.transforms as transforms
from PIL import Image

from utils import image_grid


def filter_dict(dict, keywords):
    """
    Returns only the keywords that are part of a dictionary

    Parameters
    ----------
    dictionary : dict
        Dictionary for filtering
    keywords : list of str
        Keywords that will be filtered

    Returns
    -------
    keywords : list of str
        List containing the keywords that are keys in dictionary
    """
    return [key for key in keywords if key in dict]


def resize_sample(sample, image_shape, image_interpolation=Image.ANTIALIAS):
    """
    Resizes a sample, which contains an input image.

    Parameters
    ----------
    sample : dict
        Dictionary with sample values (output from a dataset's __getitem__ method)
    shape : tuple (H,W)
        Output shape
    image_interpolation : int
        Interpolation mode

    Returns
    -------
    sample : dict
        Resized sample
    """
    # image
    image_transform = transforms.Resize(image_shape, interpolation=image_interpolation)
    sample['image'] = image_transform(sample['image'])
    return sample

def spatial_augment_sample(sample):
    """ Apply spatial augmentation to an image (flipping and random affine transformation)."""
    augment_image = transforms.Compose([
        transforms.RandomVerticalFlip(p=0.5),
        transforms.RandomHorizontalFlip(p=0.5),
        transforms.RandomAffine(15, translate=(0.1, 0.1), scale=(0.9, 1.1))
        
    ])
    sample['image'] = augment_image(sample['image'])

    return sample

def unnormalize_image(tensor, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)):
    """ Counterpart method of torchvision.transforms.Normalize."""
    for t, m, s in zip(tensor, mean, std):
        t.div_(1 / s).sub_(-m)
    return tensor


def sample_homography(
        shape, perspective=True, scaling=True, rotation=True, translation=True,
        n_scales=100, n_angles=100, scaling_amplitude=0.1, perspective_amplitude=0.4,
        patch_ratio=0.8, max_angle=pi/4):
    """ Sample a random homography that includes perspective, scale, translation and rotation operations."""

    width = float(shape[1])
    hw_ratio = float(shape[0]) / float(shape[1])

    pts1 = np.stack([[-1., -1.], [-1., 1.], [1., -1.], [1., 1.]], axis=0)
    pts2 = pts1.copy() * patch_ratio
    pts2[:,1] *= hw_ratio

    if perspective:

        perspective_amplitude_x = np.random.normal(0., perspective_amplitude/2, (2))
        perspective_amplitude_y = np.random.normal(0., hw_ratio * perspective_amplitude/2, (2))

        perspective_amplitude_x = np.clip(perspective_amplitude_x, -perspective_amplitude/2, perspective_amplitude/2)
        perspective_amplitude_y = np.clip(perspective_amplitude_y, hw_ratio * -perspective_amplitude/2, hw_ratio * perspective_amplitude/2)

        pts2[0,0] -= perspective_amplitude_x[1]
        pts2[0,1] -= perspective_amplitude_y[1]

        pts2[1,0] -= perspective_amplitude_x[0]
        pts2[1,1] += perspective_amplitude_y[1]

        pts2[2,0] += perspective_amplitude_x[1]
        pts2[2,1] -= perspective_amplitude_y[0]

        pts2[3,0] += perspective_amplitude_x[0]
        pts2[3,1] += perspective_amplitude_y[0]

    if scaling:

        random_scales = np.random.normal(1, scaling_amplitude/2, (n_scales))
        random_scales = np.clip(random_scales, 1-scaling_amplitude/2, 1+scaling_amplitude/2)

        scales = np.concatenate([[1.], random_scales], 0)
        center = np.mean(pts2, axis=0, keepdims=True)
        scaled = np.expand_dims(pts2 - center, axis=0) * np.expand_dims(
                np.expand_dims(scales, 1), 1) + center
        valid = np.arange(n_scales)  # all scales are valid except scale=1
        idx = valid[np.random.randint(valid.shape[0])]
        pts2 = scaled[idx]

    if translation:
        t_min, t_max = np.min(pts2 - [-1., -hw_ratio], axis=0), np.min([1., hw_ratio] - pts2, axis=0)
        pts2 += np.expand_dims(np.stack([np.random.uniform(-t_min[0], t_max[0]),
                                         np.random.uniform(-t_min[1], t_max[1])]),
                               axis=0)

    if rotation:
        angles = np.linspace(-max_angle, max_angle, n_angles)
        angles = np.concatenate([[0.], angles], axis=0) 

        center = np.mean(pts2, axis=0, keepdims=True)
        rot_mat = np.reshape(np.stack([np.cos(angles), -np.sin(angles), np.sin(angles),
                                       np.cos(angles)], axis=1), [-1, 2, 2])
        rotated = np.matmul(
                np.tile(np.expand_dims(pts2 - center, axis=0), [n_angles+1, 1, 1]),
                rot_mat) + center

        valid = np.where(np.all((rotated >= [-1.,-hw_ratio]) & (rotated < [1.,hw_ratio]),
                                        axis=(1, 2)))[0]

        idx = valid[np.random.randint(valid.shape[0])]
        pts2 = rotated[idx]

    pts2[:,1] /= hw_ratio

    def ax(p, q): return [p[0], p[1], 1, 0, 0, 0, -p[0] * q[0], -p[1] * q[0]]
    def ay(p, q): return [0, 0, 0, p[0], p[1], 1, -p[0] * q[1], -p[1] * q[1]]

    a_mat = np.stack([f(pts1[i], pts2[i]) for i in range(4) for f in (ax, ay)], axis=0)
    p_mat = np.transpose(np.stack(
        [[pts2[i][j] for i in range(4) for j in range(2)]], axis=0))

    homography = np.matmul(np.linalg.pinv(a_mat), p_mat).squeeze()
    homography = np.concatenate([homography, [1.]]).reshape(3,3)
    return homography

def warp_homography(sources, homography):
    """Warp features given a homography

    Parameters
    ----------
    sources: torch.tensor (1,H,W,2)
        Keypoint vector.
    homography: torch.Tensor (3,3)
        Homography.

    Returns
    -------
    warped_sources: torch.tensor (1,H,W,2)
        Warped feature vector.
    """
    _, H, W, _ = sources.shape
    warped_sources = sources.clone().squeeze()
    warped_sources = warped_sources.view(-1,2)
    warped_sources = torch.addmm(homography[:,2], warped_sources, homography[:,:2].t())
    warped_sources.mul_(1/warped_sources[:,2].unsqueeze(1))
    warped_sources = warped_sources[:,:2].contiguous().view(1,H,W,2)
    return warped_sources

def add_noise(img, mode="gaussian", percent=0.02):
    """Add image noise

    Parameters
    ----------
    image : np.array
        Input image
    mode: str
        Type of noise, from ['gaussian','salt','pepper','s&p']
    percent: float
        Percentage image points to add noise to.
    Returns
    -------
    image : np.array
        Image plus noise.
    """
    original_dtype = img.dtype
    if mode == "gaussian":
        mean = 0
        var = 0.1
        sigma = var * 0.5

        if img.ndim == 2:
            h, w = img.shape
            gauss = np.random.normal(mean, sigma, (h, w))
        else:
            h, w, c = img.shape
            gauss = np.random.normal(mean, sigma, (h, w, c))

        if img.dtype not in [np.float32, np.float64]:
            gauss = gauss * np.iinfo(img.dtype).max
            img = np.clip(img.astype(np.float) + gauss, 0, np.iinfo(img.dtype).max)
        else:
            img = np.clip(img.astype(np.float) + gauss, 0, 1)

    elif mode == "salt":
        print(img.dtype)
        s_vs_p = 1
        num_salt = np.ceil(percent * img.size * s_vs_p)
        coords = tuple([np.random.randint(0, i - 1, int(num_salt)) for i in img.shape])

        if img.dtype in [np.float32, np.float64]:
            img[coords] = 1
        else:
            img[coords] = np.iinfo(img.dtype).max
            print(img.dtype)
    elif mode == "pepper":
        s_vs_p = 0
        num_pepper = np.ceil(percent * img.size * (1.0 - s_vs_p))
        coords = tuple(
            [np.random.randint(0, i - 1, int(num_pepper)) for i in img.shape]
        )
        img[coords] = 0

    elif mode == "s&p":
        s_vs_p = 0.5

        # Salt mode
        num_salt = np.ceil(percent * img.size * s_vs_p)
        coords = tuple([np.random.randint(0, i - 1, int(num_salt)) for i in img.shape])
        if img.dtype in [np.float32, np.float64]:
            img[coords] = 1
        else:
            img[coords] = np.iinfo(img.dtype).max

        # Pepper mode
        num_pepper = np.ceil(percent * img.size * (1.0 - s_vs_p))
        coords = tuple(
            [np.random.randint(0, i - 1, int(num_pepper)) for i in img.shape]
        )
        img[coords] = 0
    else:
        raise ValueError("not support mode for {}".format(mode))

    noisy = img.astype(original_dtype)
    return noisy


def non_spatial_augmentation(img_warp_ori, jitter_paramters, color_order=[0,1,2], to_gray=False):
    """ Apply non-spatial augmentation to an image (jittering, color swap, convert to gray scale, Gaussian blur)."""

    brightness, contrast, saturation, hue = jitter_paramters
    color_augmentation = transforms.ColorJitter(brightness, contrast, saturation, hue)
    '''
    augment_image = color_augmentation.get_params(brightness=[max(0, 1 - brightness), 1 + brightness],
                                                    contrast=[max(0, 1 - contrast), 1 + contrast],
                                                    saturation=[max(0, 1 - saturation), 1 + saturation],
                                                    hue=[-hue, hue])
    '''

    B = img_warp_ori.shape[0]
    img_warp = []
    kernel_sizes = [0,1,3,5]
    for b in range(B):
        img_warp_sub = img_warp_ori[b].cpu()
        img_warp_sub = torchvision.transforms.functional.to_pil_image(img_warp_sub)

        img_warp_sub_np = np.array(img_warp_sub) 
        img_warp_sub_np = img_warp_sub_np[:,:,color_order]
        
        if np.random.rand() > 0.5:
            img_warp_sub_np = add_noise(img_warp_sub_np)

        rand_index = np.random.randint(4)
        kernel_size = kernel_sizes[rand_index]
        if kernel_size >0:
            img_warp_sub_np = cv2.GaussianBlur(img_warp_sub_np, (kernel_size, kernel_size), sigmaX=0)
        
        if to_gray:
            img_warp_sub_np = cv2.cvtColor(img_warp_sub_np, cv2.COLOR_RGB2GRAY)
            img_warp_sub_np = cv2.cvtColor(img_warp_sub_np, cv2.COLOR_GRAY2RGB)

        img_warp_sub = Image.fromarray(img_warp_sub_np)
        img_warp_sub = color_augmentation(img_warp_sub)

        img_warp_sub = torchvision.transforms.functional.to_tensor(img_warp_sub).to(img_warp_ori.device)

        img_warp.append(img_warp_sub)

    img_warp = torch.stack(img_warp, dim=0)
    return img_warp

def ha_augment_sample(data, jitter_paramters=[0.5, 0.5, 0.2, 0.05], patch_ratio=0.7, scaling_amplitude=0.2, max_angle=pi/4):
    """Apply Homography Adaptation image augmentation."""
    input_img = data['image'].unsqueeze(0)
    _, _, H, W = input_img.shape
    device = input_img.device
    
    homography = torch.from_numpy(
        sample_homography([H, W], 
        patch_ratio=patch_ratio, 
        scaling_amplitude=scaling_amplitude, 
        max_angle=max_angle)).float().to(device)
    homography_inv = torch.inverse(homography)

    source = image_grid(1, H, W,
                    dtype=input_img.dtype,
                    device=device,
                    ones=False, normalized=True).clone().permute(0, 2, 3, 1)

    target_warped = warp_homography(source, homography)
    img_warp = torch.nn.functional.grid_sample(input_img, target_warped)

    color_order = [0,1,2]
    if np.random.rand() > 0.5:
        random.shuffle(color_order)

    to_gray = False
    if np.random.rand() > 0.5:
        to_gray = True

    input_img = non_spatial_augmentation(input_img, jitter_paramters=jitter_paramters, color_order=color_order, to_gray=to_gray)
    img_warp = non_spatial_augmentation(img_warp, jitter_paramters=jitter_paramters, color_order=color_order, to_gray=to_gray)

    data['image'] = input_img.squeeze()
    data['image_aug'] = img_warp.squeeze()
    data['homography'] = homography
    data['homography_inv'] = homography_inv
    return data