missing thirdparty, cannot use matchers like lightglue/superglue

[pablovela5620]

Solution to issue cannot be found in the documentation.

• I checked the documentation.

Issue

Usually when I install hloc from source to use it I have to do a recursive clone to get things like the superpoint/superglue/lightglue network

File "/home/pablo/0Dev/personal/forked-repos/nerfstudio/.pixi/envs/default/lib/python3.10/site-packages/hloc/matchers/lightglue.py", line 2, in
from lightglue import LightGlue as LightGlue_
ModuleNotFoundError: No module named 'lightglue'

it seems like the third-party folder is missing.

Installed packages

I'm using pixi (which uses conda under the hood). this is what `pixi list` looks like

[dependencies]
python = "3.10.*"
pip = ">=24.0,<25"
cuda = {version = "*", channel="nvidia/label/cuda-11.8.0"}
pytorch-cuda = {version = "11.8.*", channel="pytorch"}
pytorch = {version = ">=2.2.0,<2.3", channel="pytorch"}
torchvision = {version = ">=0.17.0,<0.18", channel="pytorch"}
pyarrow = ">=15.0.2,<15.1"
colmap = ">=3.9.1,<3.10"
hloc = ">=1.4,<2"

### Environment info

```shell
pixi info

(nerfstudio) pablo@pablo-ubuntu:~/0Dev/personal/forked-repos/nerfstudio$ pixi info
      Pixi version: 0.19.1
          Platform: linux-64
  Virtual packages: __unix=0=0
                  : __linux=6.5.0=0
                  : __glibc=2.35=0
                  : __cuda=12.2=0
                  : __archspec=1=zen
         Cache dir: /home/pablo/.cache/rattler/cache
      Auth storage: /home/pablo/.rattler/credentials.json

Project
------------
           Version: 1.0.3
     Manifest file: /home/pablo/0Dev/personal/forked-repos/nerfstudio/pixi.toml
  Config locations: 
      Last updated: 06-05-2024 19:19:20

Environments
------------
       Environment: default
          Features: default
          Channels: nvidia/label/cuda-11.8.0, nvidia, conda-forge, pytorch
  Dependency count: 9
      Dependencies: python, pip, cuda, pytorch-cuda, pytorch, torchvision, pyarrow, colmap, hloc
 PyPI Dependencies: nerfstudio
  Target platforms: linux-64
             Tasks: tcnn-install

[Tobias-Fischer]

In conda-forge, we shouldn't add third-party dependencies. Instead, we should package them separately and add them as dependencies. You can open an issue in https://github.com/conda-forge/staged-recipes or better add a PR that adds lightglue in that repo :)

[pablovela5620]

Thank you for the quick reply! Let me see if I fully understand, In order for me to use the reconstruction/keypoint detection/matching functions in hloc with light glue via conda (assuming only lightglue and not any other third-party matcher/extractor dependencies) I would need to

Lightglue PR

1. Replace the current lightglue @ git+https://github.com/cvg/LightGlue in the requirements.txt with a packaged version (I think this would also mean I would need to get the authors of lightglue to create a pip wheel that is available via pypi?) or potentially with the kornia implementation since kornia is already a dep
2. Have the PR accepted + get a version bump of hloc so that the condafeedstock can get updated?

I'm fairly new to the packaging world outside of just being a user, so apologies if I'm asking the wrong questions!

[Tobias-Fischer]

Hi @pablovela5620 - sorry, I completely forgot to get back to you about this.

1.1 Yes, correct, ideally we would need to package lightglue here on conda-forge. This will not necessarily need a pypi release, but a release on GitHub. I created an issue: cvg/LightGlue#129
1.2 I am not sure what you mean with the kornia implementation. Do they include lightglue? kornia is already packaged on conda-forge, that would simplify things.
2. That's correct

I'm happy to help/provide input, but currently don't have the capacity to take the lead on this.

Best, Tobi

[pablovela5620]

I'd be happy to lead this, once there's a github release for lightglue I'll make a PR to use the Kornia lightglue matcher (they do indeed have lightglue already included) https://kornia.readthedocs.io/en/stable/feature.html#kornia.feature.LightGlueMatcher

[Tobias-Fischer]

Hi @pablovela5620 - I've had a closer look at this. kornia implements its own version of LightGlue, which is independent (but based on) the original authors' code. I've also started a pull request to get the original LightGlue to conda-forge: conda-forge/staged-recipes#27052

[Tobias-Fischer]

I've added a dependency to LightGlue in #4

The other third_party things are not trivial to add; if you need them, you can follow similar steps as for LightGlue.

[pablovela5620]

Awesome, Thank you for the help!