[Feature request] Common stereo metrics.

🚀 Feature

To make kornia attractive to stereo researcher, we should consider adding a set of commonly used stereo metrics as is seen in the KITTI or Middlebury benchmarks.

Metrics to consider to add:

• RootMeanSquaredError
• MeanAbsoluteError
• MeanBad{n}Error

We should consider whether we want the metrics to inherit from torchmetrics, such that the metrics can be used in Lightning modules.

Motivation

Most new stereo research all implement their own metrics - this is a good opportunity for Kornia to get a foothold in stereo research.

Pitch

In kornia.utils.metrics add above mentioned metrics. Discuss whether we want to inherit from torchmetric. The metric logic should ideally be split into functions and then referenced in a class that inherits from torchmetric in case we go that way.

Alternatives

N/A

Additional context

N/A

[shijianjian]

@shijianjian : What is the motivation behind avoiding DL nomenclature?

Not a strong objection. The term of “loss” in DL is more or less ready to be called with loss.backward(). If we are going for losses, we’d better have some support/tutorials/examples for that. Also, I think naming like kornia.measures would attract some DL libraries like torchmetrics to use us with less burden (that we are aiming on different directions).

[edgarriba]

our recent policy is just to have our (kornia non-profit org) libs dependencies except for fundamental tensor backend for now.