findeton/deepview

An implementation of Google's Deepview paper https://augmentedperception.github.io/deepview/

deepview's Issues

Colaboration

Hi!
I am also interested in the implimentaiton of the Deepview paper. Maybe we can colaborate on this. What do you think? What is the status of implimentation?

upsampling op?

Hi!
I see you are using upsampling operator
(torch.nn.Upsample(scale_factor=2,mode='bilinear') after the “Convert to RGBA” block. (LGD iteration 1 after the initialization step). Don't see this part in the paper.
Do you think they have forget to mention about it in the diagram?

A bug in utils_dset.py?

Why is the line 122 of utils_dset.py written like below:
in_pixel_y = torch.clamp(in_pixel_xy[:,1], min = crop_y / 2, max = img_h - img_h / 2)
But not:
in_pixel_y = torch.clamp(in_pixel_xy[:,1], min = crop_y / 2, max = img_h - crop_y / 2)

Because the line 121 of utils_dset.py uses "crop_x" instead of "img_w":
in_pixel_x = torch.clamp(in_pixel_xy[:,0], min = crop_x / 2, max = img_w - crop_x / 2)

I mean why for "in_pixel_y", the script uses "img_h" instead of "crop_y"?
Looking forward to your reply!