File size: 5,768 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
import torch
import torch.nn as nn
import torch.nn.functional as F

from utils import image_grid

class ConvBlock(nn.Module):
    def __init__(self, in_channels, out_channels):
        super(ConvBlock, self).__init__()
        
        self.conv = nn.Sequential(
            nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(out_channels),
            nn.ReLU(inplace=True),
            nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False),
            nn.BatchNorm2d(out_channels),
            nn.ReLU(inplace=True)
        )
    
    def forward(self, x):
        return self.conv(x)

       
class DilationConv3x3(nn.Module):
    def __init__(self, in_channels, out_channels):
        super(DilationConv3x3, self).__init__()
        
        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=2, dilation=2, bias=False)
        self.bn = nn.BatchNorm2d(out_channels)
    
    def forward(self, x):
        x = self.conv(x)
        x = self.bn(x)
        return x


class InterestPointModule(nn.Module):
    def __init__(self, is_test=False):
        super(InterestPointModule, self).__init__()
        self.is_test = is_test
        
        self.conv1 = ConvBlock(3, 32)
        self.conv2 = ConvBlock(32, 64)
        self.conv3 = ConvBlock(64, 128)
        self.conv4 = ConvBlock(128, 256)
        
        self.maxpool2x2 = nn.MaxPool2d(2, 2)
        
        # score head
        self.score_conv = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=False)
        self.score_norm = nn.BatchNorm2d(256)
        self.score_out = nn.Conv2d(256, 3, kernel_size=3, stride=1, padding=1)
        self.softmax = nn.Softmax(dim=1)
        
        # location head
        self.loc_conv = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=False)
        self.loc_norm = nn.BatchNorm2d(256)
        self.loc_out = nn.Conv2d(256, 2, kernel_size=3, stride=1, padding=1)

        # descriptor out
        self.des_conv2 = DilationConv3x3(64, 256)
        self.des_conv3 = DilationConv3x3(128, 256)

        # cross_head:
        self.shift_out = nn.Conv2d(256, 1, kernel_size=3, stride=1, padding=1)
                    
        self.relu = nn.ReLU(inplace=True)
        
    def forward(self, x):
        B, _, H, W = x.shape

        x = self.conv1(x)
        x = self.maxpool2x2(x)
        x2 = self.conv2(x)
        x = self.maxpool2x2(x2)
        x3 = self.conv3(x)
        x = self.maxpool2x2(x3)
        x = self.conv4(x)

        B, _, Hc, Wc = x.shape
        
        # score head
        score_x = self.score_out(self.relu(self.score_norm(self.score_conv(x))))
        aware = self.softmax(score_x[:, 0:2, :, :])
        score = score_x[:, 2, :, :].unsqueeze(1).sigmoid()
        
        border_mask = torch.ones(B, Hc, Wc)
        border_mask[:, 0] = 0
        border_mask[:, Hc - 1] = 0
        border_mask[:, :, 0] = 0
        border_mask[:, :, Wc - 1] = 0
        border_mask = border_mask.unsqueeze(1)
        score = score * border_mask.to(score.device)
        
        # location head        
        coord_x = self.relu(self.loc_norm(self.loc_conv(x)))
        coord_cell = self.loc_out(coord_x).tanh()
        
        shift_ratio = self.shift_out(coord_x).sigmoid() * 2.0

        step = ((H/Hc)-1) / 2.
        center_base = image_grid(B, Hc, Wc,
                                 dtype=coord_cell.dtype,
                                 device=coord_cell.device,
                                 ones=False, normalized=False).mul(H/Hc) + step

        coord_un = center_base.add(coord_cell.mul(shift_ratio * step))
        coord = coord_un.clone()
        coord[:, 0] = torch.clamp(coord_un[:, 0], min=0, max=W-1)
        coord[:, 1] = torch.clamp(coord_un[:, 1], min=0, max=H-1)

        # descriptor block
        desc_block = []
        desc_block.append(self.des_conv2(x2))
        desc_block.append(self.des_conv3(x3))
        desc_block.append(aware)

        if self.is_test:
            coord_norm = coord[:, :2].clone()
            coord_norm[:, 0] = (coord_norm[:, 0] / (float(W-1)/2.)) - 1.
            coord_norm[:, 1] = (coord_norm[:, 1] / (float(H-1)/2.)) - 1.
            coord_norm = coord_norm.permute(0, 2, 3, 1)

            desc2 = torch.nn.functional.grid_sample(desc_block[0], coord_norm)         
            desc3 = torch.nn.functional.grid_sample(desc_block[1], coord_norm)
            aware = desc_block[2]
            
            desc = torch.mul(desc2, aware[:, 0, :, :]) + torch.mul(desc3, aware[:, 1, :, :])         
            desc = desc.div(torch.unsqueeze(torch.norm(desc, p=2, dim=1), 1))  # Divide by norm to normalize.

            return score, coord, desc

        return score, coord, desc_block


class CorrespondenceModule(nn.Module):
    def __init__(self, match_type='dual_softmax'):
        super(CorrespondenceModule, self).__init__()
        self.match_type = match_type

        if self.match_type == 'dual_softmax':
            self.temperature = 0.1
        else:
            raise NotImplementedError()
 
    def forward(self, source_desc, target_desc):
        b, c, h, w = source_desc.size()       
     
        source_desc = source_desc.div(torch.unsqueeze(torch.norm(source_desc, p=2, dim=1), 1)).view(b, -1, h*w)
        target_desc = target_desc.div(torch.unsqueeze(torch.norm(target_desc, p=2, dim=1), 1)).view(b, -1, h*w)

        if self.match_type == 'dual_softmax':
            sim_mat = torch.einsum("bcm, bcn -> bmn", source_desc, target_desc) / self.temperature
            confidence_matrix = F.softmax(sim_mat, 1) * F.softmax(sim_mat, 2)
        else:
            raise NotImplementedError()
        
        return confidence_matrix