search

jahaniam / semiDepth

Tensorflow implementation of Semi-Supervised Monocular Depth Estimation with Left-Right Consistency Using Deep Neural Network.
Other
90 stars 11 forks source link

2 forward passes for left and right inverse depth computation and one backward pass ? #12

Open mitra052 opened 4 years ago

mitra052 commented 4 years ago

Hi Ali, In the code, there are two forward passes , one with left images as input and the other as right image. https://github.com/jahaniam/semiDepth/blob/f8d342dab3e370a0d7ab2721f5b4631f7b63ae40/monodepth_model.py#L508 However the gradients have been updated once as shown here: https://github.com/jahaniam/semiDepth/blob/f8d342dab3e370a0d7ab2721f5b4631f7b63ae40/monodepth_model.py#L508

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 commented 4 years ago

Apologies, the second link should have been, https://github.com/jahaniam/semiDepth/blob/f8d342dab3e370a0d7ab2721f5b4631f7b63ae40/monodepth_main.py#L181

jahaniam commented 4 years ago

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