Fine-tuning metric depth model

[ysz-leo]

The relative depth model outputs the inverse depth, while the metric depth model outputs the true depth.So I think it is necessary to use inverse depth to fine-tuning metric depth model, right?

[tmagcaya]

Maybe this is why the metric depth training gets worse over time during fine tuning

[xiaodongww]

Which loss should we use if the inverse depth is used? Should we use SiLogLoss or L2 loss? @tmagcaya @ysz-leo