Uncertainty to confidence score

[orkrukx]

Hello,
I am looking to convert the uncertainty values output by NLF into a confidence score in the range [0, 1].

I have seen how the NLF implementation computes per-frame weights that give more trust to low-uncertainty predictions for SMPL fitting:

nlf/nlf/pt/multiperson/multiperson_model.py

Lines 242 to 245 in f8611fc

	vertex_weights = vertex_uncertainties_flat**-1.5
	vertex_weights = vertex_weights / torch.mean(vertex_weights, dim=-1, keepdim=True)
	joint_weights = joint_uncertainties_flat**-1.5
	joint_weights = joint_weights / torch.mean(joint_weights, dim=-1, keepdim=True)

For now I simply compute the confidence as clamp(1 - uncertainty, 0, 1) but I'm not sure this is mathematically sound.
What would you recommend?

Thanks