TiagoCortinhal / SalsaNext

Uncertainty-aware Semantic Segmentation of LiDAR Point Clouds for Autonomous Driving
MIT License
417 stars 102 forks source link

Question about the inference code in user.py #34

Closed godspeed1989 closed 4 years ago

godspeed1989 commented 4 years ago

Thanks for your great work. I had a question in inference code from user.py. The order of model seems to be (mean, var) But in user.py: https://github.com/Halmstad-University/SalsaNext/blob/71114947a019c4760d52a4717cac531de12c67a2/train/tasks/semantic/modules/user.py#L157 And the proj_output2 is used to get classification result: https://github.com/Halmstad-University/SalsaNext/blob/71114947a019c4760d52a4717cac531de12c67a2/train/tasks/semantic/modules/user.py#L159

I am looking forward to your reply!

maxim0815 commented 4 years ago

I do have the same question. Within training you are doing it the right order:

output_mean, output_var = adf.Softmax(dim=1, keep_variance_fn=keep_variance_fn)(*output)
TiagoCortinhal commented 4 years ago

Hello @godspeed1989, @maxim0815!

You are correct it should be (mu,var). We tested with different possibilities to estimate the uncertainty and I must have forgotten to push the right order on the inference.

I fix this issue, but I will have a deeper look as soon I have the time!