File size: 7,789 Bytes
f9cbc98
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import numpy as np
from PIL import Image

import torch
from torch.utils.data import Dataset
from torchvision.transforms import ToTensor


def get_mgrid(sidelen, vmin=-1, vmax=1):
    if type(vmin) is not list:
        vmin = [vmin for _ in range(len(sidelen))]
    if type(vmax) is not list:
        vmax = [vmax for _ in range(len(sidelen))]
    tensors = tuple([torch.linspace(vmin[i], vmax[i], steps=sidelen[i]) for i in range(len(sidelen))])
    mgrid = torch.stack(torch.meshgrid(*tensors), dim=-1)
    mgrid = mgrid.reshape(-1, len(sidelen))
    return mgrid

def apply_homography(x, h):
    h = torch.cat([h, torch.ones_like(h[:, [0]])], -1)
    h = h.view(-1, 3, 3)
    x = torch.cat([x, torch.ones_like(x[:, 0]).unsqueeze(-1)], -1).unsqueeze(-1)
    o = torch.bmm(h, x).squeeze(-1)
    o = o[:, :-1] / o[:, [-1]]
    return o

def jacobian(y, x):
    B, N = y.shape
    jacobian = list()
    for i in range(N):
        v = torch.zeros_like(y)
        v[:, i] = 1.
        dy_i_dx = torch.autograd.grad(y,
                                      x,
                                      grad_outputs=v,
                                      retain_graph=True,
                                      create_graph=True)[0]  # shape [B, N]
        jacobian.append(dy_i_dx)
    jacobian = torch.stack(jacobian, dim=1).requires_grad_()
    return jacobian

def overlap_mix(img1, img2, img_order, overlap_num):
    w1 = np.linspace(0, 1, overlap_num)[::-1]
    w2 = 1 - w1
    return w1[img_order] * img1 + w2[img_order] * img2


class VideoFitting(Dataset):
    def __init__(self, path, transform=None):
        super().__init__()

        self.path = path
        if transform is None:
            self.transform = ToTensor()
        else:
            self.transform = transform

        self.video = self.get_video_tensor()
        self.num_frames, _, self.H, self.W = self.video.size()
        self.pixels = self.video.permute(2, 3, 0, 1).contiguous().view(-1, 3)
        self.coords = get_mgrid([self.H, self.W, self.num_frames])

        shuffle = torch.randperm(len(self.pixels))
        self.pixels = self.pixels[shuffle]
        self.coords = self.coords[shuffle]

    def get_video_tensor(self):
        frames = sorted(os.listdir(self.path))
        video = []
        for i in range(len(frames)):
            img = Image.open(os.path.join(self.path, frames[i]))
            img = self.transform(img)
            video.append(img)
        return torch.stack(video, 0)

    def __len__(self):
        return 1

    def __getitem__(self, idx):    
        if idx > 0: raise IndexError
            
        return self.coords, self.pixels
    

class TestVideoFitting(Dataset):
    def __init__(self, path, transform=None):
        super().__init__()

        self.path = path
        if transform is None:
            self.transform = ToTensor()
        else:
            self.transform = transform

        self.video = self.get_video_tensor()
        self.num_frames, _, self.H, self.W = self.video.size()
        self.pixels = self.video.permute(2, 3, 0, 1).contiguous().view(-1, 3)
        self.coords = get_mgrid([self.H, self.W, self.num_frames])

    def get_video_tensor(self):
        frames = sorted(os.listdir(self.path))
        video = []
        for i in range(len(frames)):
            img = Image.open(os.path.join(self.path, frames[i]))
            img = self.transform(img)
            video.append(img)
        return torch.stack(video, 0)

    def __len__(self):
        return 1

    def __getitem__(self, idx):    
        if idx > 0: raise IndexError
            
        return self.coords, self.pixels
    

class GroupVideoFitting(Dataset):
    def __init__(self, path, mask_path, transform=None, mask_transform=None):
        super().__init__()

        self.path = path
        self.mask_path = mask_path
        if transform is None:
            self.transform = ToTensor()
        else:
            self.transform = transform
        
        if mask_transform is None:
            self.mask_transform = ToTensor()
        else:
            self.mask_transform = mask_transform

        self.video = self.get_video_tensor()
        self.mask = self.get_mask_tensor()
        self.num_frames, _, self.H, self.W = self.video.size()
        self.pixels = self.video.permute(2, 3, 0, 1).contiguous().view(-1, 3)
        self.mask_pixels = self.mask.permute(2, 3, 0, 1).contiguous().view(-1, 1)
        self.coords = get_mgrid([self.H, self.W, self.num_frames])

        shuffle = torch.randperm(len(self.pixels))
        self.pixels = self.pixels[shuffle]
        self.coords = self.coords[shuffle]
        self.mask_pixels = self.mask_pixels[shuffle]

    def get_video_tensor(self):
        frames = sorted(os.listdir(self.path))
        video = []
        for i in range(len(frames)):
            img = Image.open(os.path.join(self.path, frames[i]))
            img = self.transform(img)
            video.append(img)
        return torch.stack(video, 0)
    
    def get_mask_tensor(self):
        masks = sorted(os.listdir(self.mask_path))
        all_mask = []
        for i in range(len(masks)):
            mask = Image.open(os.path.join(self.mask_path, masks[i]))
            mask = self.mask_transform(mask)
            all_mask.append(mask)
        return torch.stack(all_mask, 0)

    def __len__(self):
        return 1

    def __getitem__(self, idx):    
        if idx > 0: raise IndexError
            
        return self.coords, self.pixels, self.mask_pixels
    

class TestGroupVideoFitting(Dataset):
    def __init__(self, path, mask_path, back_mask_path, transform=None, mask_transform=None):
        super().__init__()

        self.path = path
        self.mask_path = mask_path
        self.back_mask_path = back_mask_path
        if transform is None:
            self.transform = ToTensor()
        else:
            self.transform = transform
        
        if mask_transform is None:
            self.mask_transform = ToTensor()
        else:
            self.mask_transform = mask_transform

        self.video = self.get_video_tensor()
        self.mask = self.get_mask_tensor()
        self.back_mask = self.get_back_mask_tensor()
        self.num_frames, _, self.H, self.W = self.video.size()
        self.pixels = self.video.permute(2, 3, 0, 1).contiguous().view(-1, 3)
        self.mask_pixels = self.mask.permute(2, 3, 0, 1).contiguous().view(-1, 1)
        self.back_mask_pixels = self.back_mask.permute(2, 3, 0, 1).contiguous().view(-1, 1)
        self.coords = get_mgrid([self.H, self.W, self.num_frames])

    def get_video_tensor(self):
        frames = sorted(os.listdir(self.path))
        video = []
        for i in range(len(frames)):
            img = Image.open(os.path.join(self.path, frames[i]))
            img = self.transform(img)
            video.append(img)
        return torch.stack(video, 0)
    
    def get_mask_tensor(self):
        masks = sorted(os.listdir(self.mask_path))
        all_mask = []
        for i in range(len(masks)):
            mask = Image.open(os.path.join(self.mask_path, masks[i]))
            mask = self.mask_transform(mask)
            all_mask.append(mask)
        return torch.stack(all_mask, 0)
    
    def get_back_mask_tensor(self):
        masks = sorted(os.listdir(self.back_mask_path))
        all_mask = []
        for i in range(len(masks)):
            mask = Image.open(os.path.join(self.back_mask_path, masks[i]))
            mask = self.mask_transform(mask)
            all_mask.append(mask)
        return torch.stack(all_mask, 0)

    def __len__(self):
        return 1

    def __getitem__(self, idx):    
        if idx > 0: raise IndexError
            
        return self.coords, self.pixels, self.mask_pixels, self.back_mask_pixels