How to evaluate quantitively when activating camera optimiser

[Cerf-Volant425]

For my custom dataset with imprecise poses, camera optimiser is needed. But during the training, nerfstudio jointly optimises the poses as the input poses, since this is only done to the training poses, the eval poses will no longer match up as well leading to a lower PSNR.

I was wondering, when keeping the pose optimization on, how to evaluate the results quantitively.

[tancik]

It's tricky, not sure there is a good answer. One option is to also optimize the poses for the eval also (though this could be viewed as cheating). Another option is to optimize the pose using half of the eval image and then use the other half to compute image metrics.

[Cerf-Volant425]

Thanks a lot for your help, but I still cannot get the second method clearly, is this referred to what your paper said as below?

Could you please elaborate a bit about this test-time optimisation procedure evaluation method? It would be really helpful.