Pipeline for custom data

[sagar-bharadwaj-ks]

I was looking to use hloc for following usecase which is not uncommon in vision-based localization:

• A 'map creator' creates a 'ground-truth map' of a place using a sequence of images.
• New users use this map to localize their own new images (without altering the data and features in the ground-truth map).

However, I think the examples provided don't directly address this usecase, as they are mostly written to evaluate hloc on datasets. As a result, the reference images and query images are tightly coupled. For example, in pipeline_Aachen.ipynb, the features for both the reference and query images are extracted together. Localizing query images wihtout updating the ground-truth data (eg. features db) requires changing the code.
While the notebook demo.ipynb achieves something to this effect, it uses exhaustive matching which is impractical for most usecases.
I have added pipeline_custom.ipynb that demonstrates how to use hloc for one's own data (with some very small changes to the hloc APIs).

If this is useful, I can add more documentation on how to generate one's own data (eg. generating colmap SfM model) if needed.

[sarlinpe]

Why is this function needed? why isn't pairs_from_retrieval.main sufficient?

[sagar-bharadwaj-ks]

You're right, pairs_from_retreival.main is sufficient. I will rewrite my example and push once I get the chance.