different results between monodepth2 script model(c++ libtorch) and original model(python pytorch)

[MRinel]

Hi, thank you for the work, and have you tested your converted torchscript model? The results of my test just using
monodepth2(FP32) were different from the results of original model, which has both encoder module and decoder module.

[TengFeiHan0]

@MRinel In the demo/, you can see the source code about how to do inference based on a converted torchscript model. I only merged the encoder and decoder part, which is a little different from the original code.

[MRinel]

@TengFeiHan0 thanks for your reply. I have got the disparity map from packnet.pt (FP32), like:

as a constract result from original monodepth2 two stage:

It looks better in detail from your model result compared with the original.and about 7fps in my test on 1080Ti. I wonder how to get the absolute distance(depth) from your model, I have tried just by STEREO_SCALE_FACTOR / disparity , but the result seems strange. In addition, which model do you use for conversion , monocular or mono+stereo?
Thank u.