Is there anyway to speed up the evaluation?

[guohengkai]

Hi,

Thank you for your awesome evaluation project. I'm trying to evaluate the local feature using your repo, but it is a bit slow to run one pass (https://github.com/cvg/pixel-perfect-sfm/blob/main/pixsfm/eval/eth3d/triangulation.py#L28). Is there any method to speed up the evaluation, like tuning some parameters or improving the implementation?

[puzzlepaint]

It is very likely that the evaluation can be sped up when spending some effort on optimizing it. The first step would probably be to profile it, at the minimum a basic check to see whether the accuracy or the completeness evaluation is the slower one for your datasets, to see where the most time could likely be gained.

For the accuracy evaluation, the cell counts in this line and the line below it could be tuned with little effort:

multi-view-evaluation/src/accuracy.cc

Line 36 in 0daa4f4

const int kCellCountInclination = 1024;

For the completeness evaluation, maybe this data structure could be tuned or exchanged with some kind of volumetric storage for faster access / growing:

multi-view-evaluation/src/completeness.cc

Line 108 in 0daa4f4

std::unordered_map<std::tuple<int, int, int>, CompletenessCell>

[guohengkai]

Thank you. I will try them.