2 forward passes for left and right inverse depth computation and one backward pass ?

[mitra052]

Hi Ali,
In the code, there are two forward passes , one with left images as input and the other as right image.

semiDepth/monodepth_model.py

Line 508 in f8d342d

self.invDepth_right_est= self.build_vgg_forward(self.right,'resnet50-forward',reuse=True)

However the gradients have been updated once as shown here:

semiDepth/monodepth_model.py

Line 508 in f8d342d

self.invDepth_right_est= self.build_vgg_forward(self.right,'resnet50-forward',reuse=True)

So, is it equivalent to the same weights and biases being multiplied separately to the left and right image to separately produce the left and right inverse depths?(during forward pass)
or are the weights and biases different for the left and right images ?

Thanks in advance!

[mitra052]

Apologies, the second link should have been,

semiDepth/monodepth_main.py

Line 181 in f8d342d

grads = opt_step.compute_gradients(loss)

[jahaniam]

Yes, the same weights and biases are being applied to both left and right images.