Giters

cleinc / bts

From Big to Small: Multi-Scale Local Planar Guidance for Monocular Depth Estimation

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Artifacts in upper part of KITTI images only in Pytorch model

pjckoch opened this issue · comments

commented

Hi,

I have tried out both the TF and PyTorch implementation with your pre-trained models.

I noticed that the artifacts in the upper part of the image, which are also mentioned in section 4.7 of your paper, only appear in the PyTorch version. The TF version does not seem to be affected.

Do you have any idea where this difference could come from? Did you use different parameters for training the PyTorch version?

Thanks in advance! See below for an example.

PyTorch:

2011_09_26_drive_0013_sync_0000000050.png
PyTorch_2011_09_26_drive_0013_sync_0000000050

TF:

2011_09_26_drive_0013_sync_0000000050.png
TF_2011_09_26_drive_0013_sync_0000000050

commented

@pjckoch Sorry for this late reply. As you may know, most of the ground truth depthmaps lack valid depth values on upper part on the images and this leads to the random behaviour on our network.