File size: 8,793 Bytes
872b0a0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
try:
    from spatial_correlation_sampler import SpatialCorrelationSampler
except ImportError:
    SpatialCorrelationSampler = None

class ConvBlock(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, dilation=1, activation=True):
        super().__init__()
        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride=stride, padding=dilation, dilation=dilation, bias=True)
        self.activation = nn.LeakyReLU(0.1, inplace=True) if activation else nn.Identity()

    def forward(self, x):
        return self.activation(self.conv(x))

class FeatureExtractor(nn.Module):
    def __init__(self):
        super().__init__()
        # conv1
        self.conv1_1 = ConvBlock(3, 16, stride=2)
        self.conv1_2 = ConvBlock(16, 16)
        # conv2
        self.conv2_1 = ConvBlock(16, 32, stride=2)
        self.conv2_2 = ConvBlock(32, 32)
        # conv3
        self.conv3_1 = ConvBlock(32, 64, stride=2)
        self.conv3_2 = ConvBlock(64, 64)
        # conv4
        self.conv4_1 = ConvBlock(64, 96, stride=2)
        self.conv4_2 = ConvBlock(96, 96)
        # conv5
        self.conv5_1 = ConvBlock(96, 128, stride=2)
        self.conv5_2 = ConvBlock(128, 128)
        # conv6
        self.conv6_1 = ConvBlock(128, 192, stride=2)
        self.conv6_2 = ConvBlock(192, 192)

    def forward(self, x):
        c1 = self.conv1_2(self.conv1_1(x))
        c2 = self.conv2_2(self.conv2_1(c1))
        c3 = self.conv3_2(self.conv3_1(c2))
        c4 = self.conv4_2(self.conv4_1(c3))
        c5 = self.conv5_2(self.conv5_1(c4))
        c6 = self.conv6_2(self.conv6_1(c5))
        return {
            'conv2_2': c2, 
            'conv3_2': c3, 
            'conv4_2': c4, 
            'conv5_2': c5, 
            'conv6_2': c6
        }

class ContextNetwork(nn.Module):
    def __init__(self, in_channels):
        super().__init__()
        self.conv1 = ConvBlock(in_channels, 128, dilation=1)
        self.conv2 = ConvBlock(128, 128, dilation=2)
        self.conv3 = ConvBlock(128, 128, dilation=4)
        self.conv4 = ConvBlock(128, 96, dilation=8)
        self.conv5 = ConvBlock(96, 64, dilation=16)
        self.conv6 = ConvBlock(64, 32, dilation=1)
        self.conv7 = ConvBlock(32, 2, dilation=1, activation=False)

    def forward(self, x, flow):
        x_in = torch.cat([x, flow], dim=1)
        out = self.conv1(x_in)
        out = self.conv2(out)
        out = self.conv3(out)
        out = self.conv4(out)
        out = self.conv5(out)
        out = self.conv6(out)
        delta_flow = self.conv7(out)
        return flow + delta_flow

class FlowEstimator(nn.Module):
    def __init__(self, ch_in, ch_out=2):
        super().__init__()
        self.conv1 = ConvBlock(ch_in, 128)
        self.conv2 = ConvBlock(128, 128)
        self.conv3 = ConvBlock(128, 96)
        self.conv4 = ConvBlock(96, 64)
        self.conv5 = ConvBlock(64, 32)
        self.conv6 = ConvBlock(32, ch_out, activation=False)

    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = self.conv3(x)
        x = self.conv4(x)
        features = self.conv5(x)
        flow = self.conv6(features)
        return flow, features

class DDFlowNet(nn.Module):
    def __init__(self, args=None):
        super().__init__()
        self.feature_extractor = FeatureExtractor()
        
        # PWC-Net style Cost Volume parameters
        self.search_range = 4
        if SpatialCorrelationSampler is not None:
             self.corr = SpatialCorrelationSampler(1, 9, 1, 0, 1) # kernel=1, max_disp=4 -> patch 9
        else:
             print("WARNING: SpatialCorrelationSampler not found. DDFlow will likely fail or require slow fallback.")
             self.corr = None

        # Estimators for levels 6, 5, 4, 3, 2
        # Input to estimator: CostVolume + Features + Flow(upsampled)
        # Channels:
        # cv_ch = (search_range*2+1)**2 = 81
        # Level 6: 192 ch. In: 81 + 192 + 0(flow? no flow) = 273
        self.est6 = FlowEstimator(81 + 192)
        self.est5 = FlowEstimator(81 + 128 + 2)
        self.est4 = FlowEstimator(81 + 96 + 2)
        self.est3 = FlowEstimator(81 + 64 + 2)
        self.est2 = FlowEstimator(81 + 32 + 2)
        
        self.context = ContextNetwork(32 + 2) # Feat from est2 + flow2

    def warp(self, x, flo):
        """
        warp an image/tensor (im2) back to im1, according to the optical flow
        x: [B, C, H, W] (im2)
        flo: [B, 2, H, W] flow
        """
        B, C, H, W = x.size()
        # mesh grid
        xx = torch.arange(0, W).view(1, -1).repeat(H, 1)
        yy = torch.arange(0, H).view(-1, 1).repeat(1, W)
        xx = xx.view(1, 1, H, W).repeat(B, 1, 1, 1)
        yy = yy.view(1, 1, H, W).repeat(B, 1, 1, 1)
        grid = torch.cat((xx, yy), 1).float().to(x.device)
        vgrid = grid + flo
        
        # scale grid to [-1,1]
        vgrid[:, 0, :, :] = 2.0 * vgrid[:, 0, :, :] / max(W - 1, 1) - 1.0
        vgrid[:, 1, :, :] = 2.0 * vgrid[:, 1, :, :] / max(H - 1, 1) - 1.0
        
        vgrid = vgrid.permute(0, 2, 3, 1)        
        output = F.grid_sample(x, vgrid, mode='bilinear', align_corners=True, padding_mode="border")
        return output

    def forward(self, x):
        # x: [B, 6, H, W]
        img1 = x[:, :3]
        img2 = x[:, 3:6]
        
        f1 = self.feature_extractor(img1)
        f2 = self.feature_extractor(img2)
        
        # Level 6
        # No initial flow
        # Correlation
        if self.corr is not None:
             out_corr6 = self.corr(f1['conv6_2'], f2['conv6_2']) # [B, 81, H, W]
             out_corr6 = out_corr6.view(out_corr6.shape[0], -1, out_corr6.shape[3], out_corr6.shape[4])
        else:
             B, _, H, W = f1['conv6_2'].shape
             out_corr6 = torch.zeros(B, 81, H, W).to(img1.device)
        
        inp6 = torch.cat([out_corr6, f1['conv6_2']], dim=1)
        flow6, _ = self.est6(inp6)
        
        # Level 5
        flow6_up = F.interpolate(flow6, scale_factor=2, mode='bilinear', align_corners=True) * 2.0
        f2_5_warp = self.warp(f2['conv5_2'], flow6_up)
        if self.corr is not None:
             out_corr5 = self.corr(f1['conv5_2'], f2_5_warp)
             out_corr5 = out_corr5.view(out_corr5.shape[0], -1, out_corr5.shape[3], out_corr5.shape[4])
        else:
             B, _, H, W = f1['conv5_2'].shape
             out_corr5 = torch.zeros(B, 81, H, W).to(img1.device)
        inp5 = torch.cat([out_corr5, f1['conv5_2'], flow6_up], dim=1)
        flow5, _ = self.est5(inp5)
        
        # Level 4
        flow5_up = F.interpolate(flow5, scale_factor=2, mode='bilinear', align_corners=True) * 2.0
        f2_4_warp = self.warp(f2['conv4_2'], flow5_up)
        if self.corr is not None:
             out_corr4 = self.corr(f1['conv4_2'], f2_4_warp)
             out_corr4 = out_corr4.view(out_corr4.shape[0], -1, out_corr4.shape[3], out_corr4.shape[4])
        else:
             B, _, H, W = f1['conv4_2'].shape
             out_corr4 = torch.zeros(B, 81, H, W).to(img1.device)
        inp4 = torch.cat([out_corr4, f1['conv4_2'], flow5_up], dim=1)
        flow4, _ = self.est4(inp4)

        # Level 3
        flow4_up = F.interpolate(flow4, scale_factor=2, mode='bilinear', align_corners=True) * 2.0
        f2_3_warp = self.warp(f2['conv3_2'], flow4_up)
        if self.corr is not None:
             out_corr3 = self.corr(f1['conv3_2'], f2_3_warp)
             out_corr3 = out_corr3.view(out_corr3.shape[0], -1, out_corr3.shape[3], out_corr3.shape[4])
        else:
             B, _, H, W = f1['conv3_2'].shape
             out_corr3 = torch.zeros(B, 81, H, W).to(img1.device)
        inp3 = torch.cat([out_corr3, f1['conv3_2'], flow4_up], dim=1)
        flow3, _ = self.est3(inp3)
        
        # Level 2
        flow3_up = F.interpolate(flow3, scale_factor=2, mode='bilinear', align_corners=True) * 2.0
        f2_2_warp = self.warp(f2['conv2_2'], flow3_up)
        if self.corr is not None:
             out_corr2 = self.corr(f1['conv2_2'], f2_2_warp)
             out_corr2 = out_corr2.view(out_corr2.shape[0], -1, out_corr2.shape[3], out_corr2.shape[4])
        else:
             B, _, H, W = f1['conv2_2'].shape
             out_corr2 = torch.zeros(B, 81, H, W).to(img1.device)
        inp2 = torch.cat([out_corr2, f1['conv2_2'], flow3_up], dim=1)
        flow2_raw, feat2 = self.est2(inp2)
        
        # Context Network
        flow2 = self.context(feat2, flow2_raw)
        
        if self.training:
            return flow2, flow3, flow4, flow5, flow6
        else:
            # Upsample to full res
            flow_full = F.interpolate(flow2, scale_factor=4, mode='bilinear', align_corners=True) * 4.0
            return flow_full