Dmytro Mishkin

@NicolasHug I understand your concerns. Regarding the dataset - it was not moved. The link you are referencing, is the earlier (worse) version of the dataset, published at CVPR2007 (as...

@NicolasHug actually, you know what? Let me try to reach the original link owners here and see if we can just respore everything first.

Unfortunately, that's how it is. You can try using it with `torch.compile` and usually after optimization, that is faster on CPU and GPU and uses less memory.

@erdemkkok are you sure it leaks memory and just not peaks its usage? Is the memory released back after `torch.cuda.empty_cache()`? The ways HardNet (and other patch descriptors) are extracted, is...

How many keypoints do you have? @erdemkkok

@erdemkkok it doesn't matter how many matched, 8000 descriptors extracted at once is a lot, so the memory consumption is what it is. I can make a PR, which adds...

@Mertalper24 yes, like this. If you can make PR, I will review and merge quickly

@erdemkkok it seems that it would be better for you to use some other descriptors then. I don't know any better way. DoGHardNet have benefits of the scale and rotation...

Hi, Which is the colmap version you use? Readme: ``` Warning: If you use the multi-view part of the benchmark, please use COLMAP 3.5. Newer versions change the behaviour of...

Given that 90 degrees do not require resampling and we already have , could we just add Random90XRotation?