search

TRI-ML / packnet-sfm

TRI-ML Monocular Depth Estimation Repository
https://tri-ml.github.io/packnet-sfm/
MIT License
1.23k stars 242 forks source link

ratios in SILog Loss #181

Closed adizhol closed 2 years ago

adizhol commented 2 years ago

Hi, why were the ratios chosen like this? Your paper "Geometric Unsupervised Domain Adaptation for Semantic Segmentation" doesn't mention anything about it.

https://github.com/TRI-ML/packnet-sfm/blob/6e3161f60e7161115813574557761edaffb1b6d1/packnet_sfm/losses/supervised_loss.py#L56

VitorGuizilini-TRI commented 2 years ago

Hi, those are the default parameters for most implementations of the SILog loss. We chose other loss weights using grid search, but never modified those from the default values. I hope that helps!

adizhol commented 2 years ago

also, you take the root of the si log diff

wellneck commented 10 months ago

Hope it's not too late, LOL, these ratios are from BTS # From Big to Small: Multi-Scale Local Planar Guidance for Monocular Depth Estimation: "we can see that it is a sum of the variance and a weighted squared mean of the error in log space. Therefore, setting a higher λ enforces more focusing on minimizing the variance of the error, and we use λ = 0.85 in this work. Also,we observe that properly scaling the range of the loss function improves convergence and the final training result."