File size: 6,500 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
import torch
import torch.nn as nn
import torchvision.transforms as tvf

from .modules import InterestPointModule, CorrespondenceModule

def warp_homography_batch(sources, homographies):
    """
    Batch warp keypoints given homographies. From https://github.com/TRI-ML/KP2D.

    Parameters
    ----------
    sources: torch.Tensor (B,H,W,C)
        Keypoints vector.
    homographies: torch.Tensor (B,3,3)
        Homographies.

    Returns
    -------
    warped_sources: torch.Tensor (B,H,W,C)
        Warped keypoints vector.
    """
    B, H, W, _ = sources.shape
    warped_sources = []
    for b in range(B):
        source = sources[b].clone()
        source = source.view(-1,2)
        '''
        [X,    [M11, M12, M13    [x,    M11*x + M12*y + M13           [M11, M12      [M13,
         Y,  =  M21, M22, M23  *  y, =  M21*x + M22*y + M23 = [x, y] * M21, M22    +  M23,
         Z]     M31, M32, M33]    1]    M31*x + M32*y + M33            M31, M32].T    M33]
        '''
        source = torch.addmm(homographies[b,:,2], source, homographies[b,:,:2].t())
        source.mul_(1/source[:,2].unsqueeze(1))
        source = source[:,:2].contiguous().view(H,W,2)
        warped_sources.append(source)
    return torch.stack(warped_sources, dim=0)
 
class PointModel(nn.Module):
    def __init__(self, is_test=True):
        super(PointModel, self).__init__()
        self.is_test = is_test
        self.interestpoint_module = InterestPointModule(is_test=self.is_test)
        self.correspondence_module = CorrespondenceModule()
        self.norm_rgb = tvf.Normalize(mean=[0.5, 0.5, 0.5], std=[0.225, 0.225, 0.225])
  
    def forward(self, *args):
        if self.is_test:
            img = args[0]
            img = self.norm_rgb(img)
            score, coord, desc = self.interestpoint_module(img)
            return score, coord, desc
        else:
            source_score, source_coord, source_desc_block = self.interestpoint_module(args[0])
            target_score, target_coord, target_desc_block = self.interestpoint_module(args[1])

            B, _, H, W = args[0].shape
            B, _, hc, wc = source_score.shape
            device = source_score.device

            # Normalize the coordinates from ([0, h], [0, w]) to ([0, 1], [0, 1]).
            source_coord_norm = source_coord.clone()
            source_coord_norm[:, 0] = (source_coord_norm[:, 0] / (float(W - 1) / 2.)) - 1.
            source_coord_norm[:, 1] = (source_coord_norm[:, 1] / (float(H - 1) / 2.)) - 1.
            source_coord_norm = source_coord_norm.permute(0, 2, 3, 1)

            target_coord_norm = target_coord.clone()
            target_coord_norm[:, 0] = (target_coord_norm[:, 0] / (float(W - 1) / 2.)) - 1.
            target_coord_norm[:, 1] = (target_coord_norm[:, 1] / (float(H - 1) / 2.)) - 1.
            target_coord_norm = target_coord_norm.permute(0, 2, 3, 1)
            
            target_coord_warped_norm = warp_homography_batch(source_coord_norm, args[2])
            target_coord_warped = target_coord_warped_norm.clone()
        
            # de-normlize the coordinates
            target_coord_warped[:, :, :, 0] = (target_coord_warped[:, :, :, 0] + 1) * (float(W - 1) / 2.)
            target_coord_warped[:, :, :, 1] = (target_coord_warped[:, :, :, 1] + 1) * (float(H - 1) / 2.)
            target_coord_warped = target_coord_warped.permute(0, 3, 1, 2)

            # Border mask
            border_mask_ori = torch.ones(B, hc, wc)
            border_mask_ori[:, 0] = 0
            border_mask_ori[:, hc - 1] = 0
            border_mask_ori[:, :, 0] = 0
            border_mask_ori[:, :, wc - 1] = 0
            border_mask_ori = border_mask_ori.gt(1e-3).to(device)

            oob_mask2 = target_coord_warped_norm[:, :, :, 0].lt(1) & target_coord_warped_norm[:, :, :, 0].gt(-1) & target_coord_warped_norm[:, :, :, 1].lt(1) & target_coord_warped_norm[:, :, :, 1].gt(-1)
            border_mask = border_mask_ori & oob_mask2

            # score
            target_score_warped = torch.nn.functional.grid_sample(target_score, target_coord_warped_norm.detach(), align_corners=False)

            # descriptor
            source_desc2 = torch.nn.functional.grid_sample(source_desc_block[0], source_coord_norm.detach())
            source_desc3 = torch.nn.functional.grid_sample(source_desc_block[1], source_coord_norm.detach())
            source_aware = source_desc_block[2]
            source_desc = torch.mul(source_desc2, source_aware[:, 0, :, :].unsqueeze(1).contiguous()) + torch.mul(source_desc3, source_aware[:, 1, :, :].unsqueeze(1).contiguous())

            target_desc2 = torch.nn.functional.grid_sample(target_desc_block[0], target_coord_norm.detach())
            target_desc3 = torch.nn.functional.grid_sample(target_desc_block[1], target_coord_norm.detach())
            target_aware = target_desc_block[2]
            target_desc = torch.mul(target_desc2, target_aware[:, 0, :, :].unsqueeze(1).contiguous()) + torch.mul(target_desc3, target_aware[:, 1, :, :].unsqueeze(1).contiguous())
            
            target_desc2_warped = torch.nn.functional.grid_sample(target_desc_block[0], target_coord_warped_norm.detach())
            target_desc3_warped = torch.nn.functional.grid_sample(target_desc_block[1], target_coord_warped_norm.detach())
            target_aware_warped = torch.nn.functional.grid_sample(target_desc_block[2], target_coord_warped_norm.detach())
            target_desc_warped = torch.mul(target_desc2_warped, target_aware_warped[:, 0, :, :].unsqueeze(1).contiguous()) + torch.mul(target_desc3_warped, target_aware_warped[:, 1, :, :].unsqueeze(1).contiguous())
            
            confidence_matrix = self.correspondence_module(source_desc, target_desc)
            confidence_matrix = torch.clamp(confidence_matrix, 1e-12, 1 - 1e-12)
            
            output = {
                'source_score': source_score,
                'source_coord': source_coord,
                'source_desc': source_desc,
                'source_aware': source_aware,
                'target_score': target_score,
                'target_coord': target_coord,
                'target_score_warped': target_score_warped,
                'target_coord_warped': target_coord_warped,
                'target_desc_warped': target_desc_warped,
                'target_aware_warped': target_aware_warped,
                'border_mask': border_mask,
                'confidence_matrix': confidence_matrix
            }
        
            return output