Fix disparity smoothness missing y gradient

[braun-steven]

In get_disparity_smoothness:

monodepth/monodepth_model.py

Line 121 in 5bc4bb1

return smoothness_x + smoothness_y

)

the two lists for the x and y direction are concatenated while the loss in your paper suggests to use elementwise addition. To fix this you could either pad the gradients on the right side and bottom by 1 or simply iterate over the 8 instead of 4 elements of the gradients list. Else you disregard the y gradients as the first 4 entries are only the x gradients.

[Songminkee]

self.disp_left_loss = [tf.reduce_mean(tf.abs(self.disp_left_smoothness[i])) / 2 ** i for i in range(4 * 2)]
self.disp_right_loss = [tf.reduce_mean(tf.abs(self.disp_right_smoothness[i])) / 2 **i for i in range(4 * 2)]

Wouldn't this be better?

self.disp_left_loss = [tf.reduce_mean(tf.abs(self.disp_left_smoothness[i])) / 2 ** (i%4) for i in range(4 * 2)]
self.disp_right_loss = [tf.reduce_mean(tf.abs(self.disp_right_smoothness[i])) / 2 **(i%4) for i in range(4 * 2)]

[asherp7]

When debugging the code I ran into this bug too. I wonder if fixing this will give better results...