Spaces:
Running
Running
File size: 6,199 Bytes
a80d6bb 472119d a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 a80d6bb c74a070 |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from lanet_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) / 2.0
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) / 2.0)) - 1.0
coord_norm[:, 1] = (coord_norm[:, 1] / (float(H - 1) / 2.0)) - 1.0
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
|