Usage on custom dataset

[EarltShirt]

Hello,
thanks a lot for your work. I wanted to know if your method was compatible with non static monocular videos as input ? I would like to test it out on a single video which turns around a moving object, is it somehow possible ?
Thanks for your time

[EarltShirt]

Hello,

Sorry to bother you again, could you explain how to the scene's extent is computed and if it is overritable ? I'm speaking of the attribute scene.cameras_extent, I would like to know if it corresponds to the size of the scene as one can see when opening the pointcloud in a external software (as cloudcompare). Lastly, I wanted to knwo if it is possible to use a specific pointcloud as a base for our gaussians: I saw that a file named points3d.ply gets generated each tim i train your model on a new scene, and wanted to know how to indicate (for example in the custom yaml configuration file) that I want to use a specific startpoint for the gaussians. Finally, I wanted to know if it was possible to slightly change the way the initialization process and the densification works: for the initialization, is it possible to change the size of the sphere on which gaussians are added to simulate a background (in the case of monocular video settings as on the lego dnerf dataset), and for the densification, is there a way to verify that the gaussians are generated at the right place (since the scene.cameras_extent attribute is used, I don't know if it corresponds to the actual value and thus if the densification process is working).

Thanks a lot for your work and your time.

[Alexander0Yang]

For the the camera extent, we inherit the implementation in original 3DGS, which is defined as the maximum distance from all cameras to the centroid of them. Specifing a custom initial point cloud just need to substitute the points3d.ply in the folder of used scene.