Jkulhanek/colmap dataparser

[jkulhanek]

This PR implements a dataparser to load the COLMAP sparse model directly without the need to run ns-process-data.

[jkulhanek]

I think this dataparser can make it easier for users to use nerfstudio as ns-process-data is no longer required. Only the COLMAP model is sufficient. Currently, ns-process-data does three things:

1. create colmap model
2. downscale images
3. export data as a json

If we accept this dataparser we can instead have a script that would only run colmap (for users which don't run it themselves).

This dataparser should be compatible with folders generated using ns-process-data (as they store the COLMAP model at the same paths). I would even go as far as suggest this as the default for nerfacto and other models. Let me know what you think.

[tancik]

LGTM

[tancik]

I think this dataparser can make it easier for users to use nerfstudio as ns-process-data is no longer required. Only the COLMAP model is sufficient. Currently, ns-process-data does three things:

1. create colmap model
2. downscale images
3. export data as a json

If we accept this dataparser we can instead have a script that would only run colmap (for users which don't run it themselves).

This dataparser should be compatible with folders generated using ns-process-data (as they store the COLMAP model at the same paths). I would even go as far as suggest this as the default for nerfacto and other models. Let me know what you think.

Wouldn't this conflict with users that use non-colmap ns-process-data pipelines? It seems like many people use metashape or realitycapture.

[jkulhanek]

I think this dataparser can make it easier for users to use nerfstudio as ns-process-data is no longer required. Only the COLMAP model is sufficient. Currently, ns-process-data does three things:

1. create colmap model
2. downscale images
3. export data as a json

If we accept this dataparser we can instead have a script that would only run colmap (for users which don't run it themselves).
This dataparser should be compatible with folders generated using ns-process-data (as they store the COLMAP model at the same paths). I would even go as far as suggest this as the default for nerfacto and other models. Let me know what you think.

Wouldn't this conflict with users that use non-colmap ns-process-data pipelines? It seems like many people use metashape or realitycapture.

Yes, let’s keep the default then. I didn’t know many people used non colmap pipelines and it would make it more difficult for them.

[hdacnw]

Found out that colmap dataparser only works with absolute path of the directory containing colmap outputs. Do you know where can I change for it to work with relative path? Thanks!