search

wvangansbeke / Sparse-Depth-Completion

Predict dense depth maps from sparse and noisy LiDAR frames guided by RGB images. (Ranked 1st place on KITTI) [MVA 2019]
https://arxiv.org/pdf/1902.05356.pdf
Other
484 stars 78 forks source link

Training with Confidence for Global and Local Networks #18

Closed saikumar96 closed 4 years ago

saikumar96 commented 4 years ago

From the paper: "We first train both parts of the framework individually and use a pretrained ERFNet on Cityscapes [22] for our global network." I was wondering how global and local networks are trained whether with or without confidence maps and how is the final end to end training done, with these pre-trained sub-networks

wvangansbeke commented 4 years ago

Hi @saikumar96,

Yes you are right. Both parts can be trained individually and later everything can be trained together. When you use skip connections between the local and global network, it's not that important. Differences are negligable. This version can be trained end-to-end without problem.

Best, Wouter

LiangliangGuo commented 4 years ago

Hi @wvangansbeke

What I want to know is whether you first train global net and local net separately, and the combine then for end to end?

wvangansbeke commented 4 years ago

Hi @LiangliangGuo,

I initially did it like you described. Now in the latest version, I simply trained it end-to-end. It should give you the same numbers.

Best, Wouter

LiangliangGuo commented 4 years ago

@wvangansbeke Thank you very much for your kind answer!!!