Speed up reconstruction using external initial pose estimates

[knicklicht]

I noticed that the reconstruction for large scenes take a long time. Would providing initial guesses for camera poses that were aquired through some external mechanism speed up the pipeline? I am thinking of inertial measurement units or maybe even from a SLAM system rigidly attached to the camera? Where in the OpenSfM pipeline would I need to set these initial guesses?

[fabianschenk]

Hi @knicklicht,

What operation takes the longest in your experiment? Is it really the reconstruction itself or the matching? Did you try to split the dataset and to run OpenSfM each subset separately as described here? If it's the matching action, then you would need to implement some strategy to find "nearby" images in order to speed up the matching. If it's the reconstruction action, then you can try to implement something that runs the incremental reconstruction but uses a prior. Maybe this can help: https://github.com/mapillary/OpenSfM/blob/a1a253f148d6e5daf6f2247e43e7d3fd306def53/opensfm/reconstruction.py#L1624

[kielnino]

I would also like to have the ability to insert prior pose informartion. Unfortunately reconstruct_from_prior is only suitable to add new images to an existing reconstruction.

[fabianschenk]

Hi @kielnino ,

This is an open-source project, so if you find something lacks functionality, feel free to create a PR :smile:

Best, Fabian

[kielnino]

@fabianschenk Sure, but then the PRs must also be accepted at some point. My other PR is waiting for a month now (https://github.com/mapillary/OpenSfM/pull/957)

Best, Johannes

[fabianschenk]

Hi @kielnino ,

Fair point 😉 I imported the PR and we'll review it. As you might have guessed from the very few code changes in the last months, the development of OpenSfM has slowed down a bit on our side.

Best, Fabian

[kielnino]

@fabianschenk thank you. I have noticed the slow down but as the project is very mature, it is a pleasure to work with anyway. Thanks for making it public available. Maybe I will follow your suggestion and give the implementation of prior pose import a try by myself.