How to measure the consistency across multiple views, rather than just between a pair of views?

[MiaApr]

No description provided.

[mohammadasim98]

Hi, thank you for taking an interest in our work.

Currently, we only consider 2 images at a time. Ideally, for multiple views, you need to compute all possible combinations, i.e., ⁿC₂ where n is the total number of images, and then take the average. However this becomes inefficient if n is larger.

In that case, if you have an ordered sequence of images, then you can perform a pairwise evaluation as done in the original paper, where you compute MEt3R in a sliding window fashion and then take the average. this I more efficient as compared to dense evaluation and proved to be sufficient for comparing different level of inconsistencies.

In case the images are not ordered, then you can define a set of fixed pairs (not necessarily all combinations, but make sure that there are overlaps between pairs) and evaluate MEt3R on these. Then, for comparing methods, use the same set of pairs for fairness.