Disentangling style and geometry

[abcp4]

There is some example code to see the Disentangling style and geometry, where one keeps the appearance of the source, but with the target geometry, as shown in the paper?

[tomasjakab]

Hi! This should be easy to do by adapting the Jupyter notebook examples/visualize.ipynb.

You just need to modify the dictionary on this line
dataset = tf.data.Dataset.from_tensors({'image': images, 'future_image': images})
The values of the keys 'image' and 'future_image' should be the appearance and pose images respectively.

Then you need to change this line
landmarks = session.run(tensors['gauss_yx'])
to something like this
reconstructed_images = session.run(tensors['future_im_pred'])
This will give you reconstructed images that have the appearance from 'image' images and pose from 'future_image' images.

Hope this helps.

(I might also provide a full script for that at some point.)

[abcp4]

It works, thanks!

[tomasjakab]

Awesome!