Question training data resolution! about sparse-to-dense.pytorch HOT 6 CLOSED

To get a size of 480-by-640, you need the following changes in your fork:

1. modify the train transform and the validation transform to load images of size 480-by-640
2. modify the output file size in the model definition

   sparse-to-dense.pytorch/models.py

   Line 9 in f16e13f

   oheight, owidth = 228, 304

Please let me know if these changes work.

from sparse-to-dense.pytorch.

What exactly is the reason for sampling down to 228x304? In your paper you cite [3] and [13] but they are also not really explaining the reason for that.

from sparse-to-dense.pytorch.

What exactly is the reason for sampling down to 228x304? In your paper you cite [3] and [13] but they are also not really explaining the reason for that.

I downsampled to 228x304 simply for the sake of comparison against previous methods. I don't think there's anything magical to these numbers. Any size that is above 224x224 (minimal size for ResNet) should work fine.

from sparse-to-dense.pytorch.

@fangchangma After I modify the input and output size to 480*640, model works well.

from sparse-to-dense.pytorch.

@fangchangma I am also trying to resize the model to train and output with a size of 480-640 (original image size of the NYU Depth data), but the implementation is now slightly different from your previous instructions since the code has changed.

I though the only necessary modification is to change the following line of code from (228, 304) to (480,640).

Unfortunately, this results in the following error:
RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0. Got 81 and 60 in dimension 2 at /pytorch/aten/src/TH/generic/THTensorMoreMath.cpp:1307

Any suggestions?

from sparse-to-dense.pytorch.

@duanyongli @RCConnolly @timethy @AbigailFernandes Hi, the download speed (13kb/s ) is too low when I use the command "wget http://datasets.lids.mit.edu/sparse-to-dense/data/nyudepthv2.tar.gz" even in VPN mode. How can I raise the download speed? Looking forward to your reply.

from sparse-to-dense.pytorch.