Open 3942368 opened 1 week ago
Hi, thanks for reporting, could you check inside the conda environment prepared for DIM if cuda is really available?
> python
> import torch
> torch.cuda.is_available()
here you are
Hi, thanks for reporting, could you check inside the conda environment prepared for DIM if cuda is really available?
> python > import torch > torch.cuda.is_available()
This is the result from running on the Colab environment set up on your side. It was a T4 GPU with the aliked+lightglue configuration. Of course, PyTorch is properly applied (since it's your Colab code). However, the speed in matching is only 1.27s/it. It seems like the GPU is not being fully utilized.
Hi, thanks for reporting. Did you solve the issue? A solution to significantly decrease the matching time is to use GeometricVerification.MAGSAC
instead of GeometricVerification.PYDEGENSAC
that is much slower
I experimented with superpoint + lightglue and aliked + lightglue.
The GPU on my local machine is a 1060 6GB, and the cloud environment on lightning.ai is using an L4 GPU.
On my local machine, the matcher speed is around 1.56 it/s, but in the cloud, it only reaches 3.52 it/s. I tested with around 600 images at a resolution of 1280x720. The command I used for testing was: -p aliked+lightglue --skip_reconstruction -s sequential --overlap 30 It seems like the matchers aren’t utilizing the GPU properly. Isn't this speed an issue?
this is local gpu speed(1060 6G) this is cloud gpu(L4)