Image comparison of the ground truth depth and predicted depth about monodepth HOT 4 CLOSED

Hi,

Your errors seem quite low, it looks like the training went ok.
I do not visualize the depth directly but the inverse of it (or disparity).

If you try to save 1.0 / pred_depth you should see a reasonable image.

from monodepth.

ok, let me try that

from monodepth.

ok, I got the disparity images which look reasonable. By the way in the function def scale_pyramid(self, img, num_scales): I had to cast the nh and nw as tensor elements otherwise the code was giving errors,

nh = tf.cast(nh, tf.int32)
nw = tf.cast(nw, tf.int32)
scaled_imgs.append(tf.image.resize_area(img, [nh, nw]))

The above change worked for me. I am using Python3 though and I guess your code is based on Python2.

I had one more question - your encoder for vgg as I see is pretty straightforward but for the decoder part is there any related article or link that I can read upon to get a better feel of it and how the disparities are getting output at different levels.

from monodepth.

The code was written for python2, I'll need to update at some point to make sure it works for both.

The decoder, is inspired by dispnet and flownet from Thomas' Brox group.
https://arxiv.org/abs/1504.06852

from monodepth.