2gunsu / monocon-pytorch

Unofficial Pytorch Implementation for MonoCon(AAAI, 2022)
Apache License 2.0
86 stars 10 forks source link

The LAUL depth loss going Negative #29

Open aminullah6264 opened 9 months ago

aminullah6264 commented 9 months ago

The Laplacian Aleatoric Uncertainty Loss accounts for uncertainty in the predictions. The idea is to weigh the loss based on the predicted uncertainty (represented by log_variance). However, the loss is going negative, is it normal or have the author observed this behavior of the depth loss?

image