def inference(self, left, right):
cost_volume = compute_cost_volume(right, left)
linear = self.linear(cost_volume)
en = self.en(linear)
up_en = self.upconv(en)
de = self.de(torch.cat((up_en, linear), dim=1))
scale_pred = self.regress(de)
return scale_pred
def inference(self, left, right):
cost_volume = compute_cost_volume(right, left)
x = self.linear(cost_volume)
# encoder
x2s = self.en1(x)
fc = self.fc(x2s)
x2s = self.de1(torch.cat([fc, x2s], dim=1))
de_x = F.interpolate(x2s, scale_factor=2, mode='bilinear')
scale_pred = self.scale_pred(torch.cat([de_x, x], dim=1))
return F.softmax(scale_pred, dim=1)
def forward(self, left, right, kps1, images1, scales):
"""
left: [B, D, H0, W0]
right: [B, D, H1, W1]
kps1: [B, N, 2], belongs to images1
images1: [B, D, H1', W1']
scales: [B, D, H1', W1']
"""
# [b, h0*w0, h1, w1]
cost_volume = compute_cost_volume(right, left)
scales_pred = self.regress(cost_volume)
scale_loss = self.scale_evaluator(scales_pred, scales, kps1, images1)
return scales_pred, scale_loss
def forward(self, left, right, scale, msk):
"""
left: [B, D, H0, W0]
right: [B, D, H1, W1]
scale: [B, D, H1', W1']
msk: [B, H1', W1']
"""
# [b, h0*w0, h1, w1]
cost_volume = compute_cost_volume(right, left)
x = self.linear(cost_volume)
# encoder
x2s = self.en1(x)
fc = self.fc(x2s)
x2s = self.de1(torch.cat([fc, x2s], dim=1))
de_x = F.interpolate(x2s, scale_factor=2, mode='bilinear')
scale_pred = self.scale_pred(torch.cat([de_x, x], dim=1))
# encoder
# x2s = self.en1(x)
# x4s = self.en2(x2s)
# fc = self.fc(x4s)
# decoder
# x4s = self.de1(torch.cat([fc, x4s], dim=1))
# de_x2s = F.interpolate(x4s, scale_factor=2, mode='bilinear')
# x2s = self.de2(torch.cat([de_x2s, x2s], dim=1))
# de_x = F.interpolate(x2s, scale_factor=2, mode='bilinear')
# scale_pred = self.scale_pred(de_x)
h1, w1 = right.shape[2:]
scale = F.interpolate(scale.unsqueeze(1), (h1, w1), mode='bilinear')
scale[scale > 1.5] = 2
scale[(scale >= 0.75) * (scale <= 1.5)] = 1
scale[scale < 0.75] = 0
scale_loss = self.scale_evaluator(scale_pred, scale.long().squeeze(1))
msk = F.interpolate(msk.unsqueeze(1), (h1, w1),
mode='bilinear').squeeze(1)
scale_loss = torch.sum(scale_loss * msk) / torch.sum(msk)
return F.softmax(scale_pred, dim=1), scale_loss
def forward(self, left, right, scale, msk):
"""
left: [B, D, H0', W0']
right: [B, D, H1', W1']
scale: [B, 1, H1, W1]
msk: [B, 1, H1, W1]
"""
# [b, h0*w0, h1, w1]
cost_volume = compute_cost_volume(right, left)
scale_pred = self.regress(cost_volume)
h1, w1 = right.shape[2:]
scale = F.interpolate(scale, (h1, w1), mode='bilinear')
scale[scale > 1.5] = 2
scale[(scale >= 0.75) * (scale <= 1.5)] = 1
scale[scale < 0.75] = 0
scale_loss = self.scale_evaluator(scale_pred, scale.long().squeeze(1))
msk = F.interpolate(msk, (h1, w1), mode='bilinear').squeeze(1)
scale_loss = torch.sum(scale_loss * msk) / torch.sum(msk)
return F.softmax(scale_pred, dim=1), scale_loss
def forward(self, left, right, scale, msk):
"""
left: [B, D, H0, W0]
right: [B, D, H1, W1]
scale: [B, D, H1', W1']
msk: [B, H1', W1']
"""
# [b, h0*w0, h1, w1]
cost_volume = compute_cost_volume(right, left)
linear = self.linear(cost_volume)
en = self.en(linear)
up_en = self.upconv(en)
de = self.de(torch.cat((up_en, linear), dim=1))
scale_pred = self.regress(de)
h1, w1 = right.shape[2:]
scale = F.interpolate(scale.unsqueeze(1), (h1, w1), mode='bilinear')
msk = F.interpolate(msk.unsqueeze(1), (h1, w1), mode='bilinear').long().float()
scale_loss = self.scale_evaluator(scale_pred, scale, msk)
return scale_pred, scale_loss
def inference(self, left, right):
cost_volume = compute_cost_volume(right, left)
scale_pred = self.regress(cost_volume)
return scale_pred