The problem of the loss function

zzangjinsun / NLSPN_ECCV20

Park et al., Non-Local Spatial Propagation Network for Depth Completion, ECCV, 2020

MIT License

321 stars 57 forks source link

The problem of the loss function #57

Closed sunohm closed 2 months ago

sunohm commented 4 months ago

Hello, Thank you for your contribution. I have a question I would like to ask you. Why use a weighted sum of L1 and L2 loss, 1.0L1+1.0L2?

zzangjinsun commented 3 months ago

We empirically found that combining L1 and L2 losses effectively balances RMSE and MAE. It is also helpful to get a better RMSE in the KITTIDC dataset.

The ablation study results in the supplementary material are given as follows:

Please refer to the supplementary material for more details.