search

TRI-ML / packnet-sfm

TRI-ML Monocular Depth Estimation Repository
https://tri-ml.github.io/packnet-sfm/
MIT License
1.21k stars 241 forks source link

About difference between PackNetSlim01.py and PackNetSAN01.py in packnet sub-module #203

Open mzy97 opened 2 years ago

mzy97 commented 2 years ago

In the issue https://github.com/TRI-ML/packnet-sfm/issues/163, you said packnet-san use packnet slim as rgb encoder decoder, but when comparing PackNetSlim01.py and PackNetSAN01.py, I found that there some change in PackNetSAN01.py that PackNetSAN01 doesn't support version "B" (adding fusion on decoder), why make this change and how the skip connection fusion method affect the accuracy? Thanks

mzy97 commented 2 years ago

And plus, I also curious about the UnpackLayerConv3d hyper-parameter settings in code https://github.com/TRI-ML/packnet-sfm/blob/f59b1d615777a9987285a10e45b5d87b0369fa7d/packnet_sfm/networks/depth/PackNetSlim01.py#L74-L79 unpack4 suddenly decrease the channel by 4 times when setting version to 'B', these design I think is not common, why you make such design. Thank you

VitorGuizilini-TRI commented 2 years ago

Sorry for the late reply. We don't use the 'B' version in PackNet-SAN, so it is probably outdated. We did some internal experiments on the original PackNet paper, and didn't see much difference between versions.