def forward(self, theta):
# shape of theta: [b, 6]
b = theta.size(0)
if not theta.size() == (b, 6):
theta = theta.view(b, 6)
theta = theta.contiguous()
# shape: [1, h, w, 1]
grid_X = torch.Tensor(self.grid_X).unsqueeze(0).unsqueeze(3).cuda()
grid_Y = torch.Tensor(self.grid_Y).unsqueeze(0).unsqueeze(3).cuda()
grid_X = Variable(grid_X, requires_grad=False)
grid_Y = Variable(grid_Y, requires_grad=False)
# shape of each theta element: [b, 1, 1, 1]
t0 = theta[:, 0].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t1 = theta[:, 1].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t2 = theta[:, 2].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t3 = theta[:, 3].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t4 = theta[:, 4].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t5 = theta[:, 5].unsqueeze(1).unsqueeze(2).unsqueeze(3)
X = expand_dim(grid_X, 0, b) # shape: [b, h, w, 1]
Y = expand_dim(grid_Y, 0, b) # shape: [b, h, w, 1]
# P = Theta @ [X, Y, 1].T
Xp = X * t0 + Y * t1 + t2 # shape: [b, h, w, 1]
Yp = X * t3 + Y * t4 + t5 # shape: [b, h, w, 1]
return torch.cat((Xp, Yp), 3) # shape: [b, h, w, 2]
def forward(self, theta, matches, return_outliers=False):
if isinstance(theta, Variable): # handle normal batch transformations
batch_size = theta.size()[0]
theta = theta.clone()
# print(self.mask_id.size())
mask = self.geometricTnf(expand_dim(self.mask_id, 0, batch_size),
theta)
# print(mask.size())
# mask = self.maxpool2d(mask)
if return_outliers:
mask_outliers = self.geometricTnf(
expand_dim(1.0 - self.mask_id, 0, batch_size), theta)
if self.normalize:
epsilon = 1e-5
mask = torch.div(
mask,
torch.sum(
torch.sum(torch.sum(mask + epsilon, 3), 2),
1).unsqueeze(1).unsqueeze(2).unsqueeze(3).expand_as(
mask)) # 有个吉儿用？mask 本来就是 0/1 矩阵，搁这 normalize 你🐎呢
if return_outliers:
mask_outliers = torch.div(
mask_outliers,
torch.sum(
torch.sum(torch.sum(mask_outliers + epsilon, 3),
2),
1).unsqueeze(1).unsqueeze(2).unsqueeze(
3).expand_as(mask_outliers))
score = torch.sum(
torch.sum(torch.sum(torch.mul(mask, matches), 3), 2), 1)
if return_outliers:
score_outliers = torch.sum(
torch.sum(torch.sum(torch.mul(mask_outliers, matches), 3),
2), 1)
return (score, score_outliers)
elif isinstance(
theta, list
): # handle multiple transformations per batch item, batch is in list format (used for RANSAC)
batch_size = len(theta)
score = []
for b in range(batch_size):
sample_size = theta[b].size(0)
s = self.forward(
theta[b],
expand_dim(matches[b, :, :, :].unsqueeze(0), 0,
sample_size))
score.append(s)
return score
def forward(self, theta):
b = theta.size(0)
if not theta.size() == (b, 6):
theta = theta.view(b, 6)
theta = theta.contiguous()
t0 = theta[:, 0].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t1 = theta[:, 1].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t2 = theta[:, 2].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t3 = theta[:, 3].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t4 = theta[:, 4].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t5 = theta[:, 5].unsqueeze(1).unsqueeze(2).unsqueeze(3)
X = expand_dim(self.grid_X, 0, b)
Y = expand_dim(self.grid_Y, 0, b)
Xp = X * t0 + Y * t1 + t2
Yp = X * t3 + Y * t4 + t5
return torch.cat((Xp, Yp), 3)
def forward(self, theta):
b = theta.size(0)
if not theta.size() == (b, 6):
theta = theta.view(b, 6)
theta = theta.contiguous()
grid_X = torch.Tensor(self.grid_X).unsqueeze(0).unsqueeze(3).cuda()
grid_Y = torch.Tensor(self.grid_Y).unsqueeze(0).unsqueeze(3).cuda()
grid_X = Variable(grid_X, requires_grad=False)
grid_Y = Variable(grid_Y, requires_grad=False)
t0 = theta[:, 0].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t1 = theta[:, 1].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t2 = theta[:, 2].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t3 = theta[:, 3].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t4 = theta[:, 4].unsqueeze(1).unsqueeze(2).unsqueeze(3)
t5 = theta[:, 5].unsqueeze(1).unsqueeze(2).unsqueeze(3)
X = expand_dim(grid_X, 0, b)
Y = expand_dim(grid_Y, 0, b)
Xp = X * t0 + Y * t1 + t2
Yp = X * t3 + Y * t4 + t5
return torch.cat((Xp, Yp), 3)