search

TiagoCortinhal / SalsaNext

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

Some inference questions about final mean and variance #41

Closed fcyeh closed 3 years ago

fcyeh commented 3 years ago

image Based on the paper "A General Framework for Uncertainty Estimation in Deep Learning", I have questions about your inference code in user.py.

  1. Should the final prediction be average of proj_output_r or proj_output2? The equation from the paper is average of T times monte-carlo predictions. https://github.com/Halmstad-University/SalsaNext/blob/cc8c75dc68d2607d16e2c82be61e7254e5b74a12/train/tasks/semantic/modules/user.py#L157

  2. Is log_var2 sensor uncertainty? If yes, based on the equation, we just need to do average. Why did you do log_var_r.var()? https://github.com/Halmstad-University/SalsaNext/blob/cc8c75dc68d2607d16e2c82be61e7254e5b74a12/train/tasks/semantic/modules/user.py#L159

Thank you.

TiagoCortinhal commented 3 years ago

Hello @fcyeh!

It should be yes, but given our output is not a linear like in their case, taking directly the mean did not work.

fcyeh commented 3 years ago

Hi @TiagoCortinhal. Thank you for the response. I think you were talking about question 1 right? How about question 2, the sensor uncertainty? Why need to do var before average? https://github.com/Halmstad-University/SalsaNext/blob/cc8c75dc68d2607d16e2c82be61e7254e5b74a12/train/tasks/semantic/modules/user.py#L159 The equation in the paper: image Thank you.

TiagoCortinhal commented 3 years ago

Yes it should be mean. This was corrected!

fcyeh commented 3 years ago

Thank you!