Add Metadata Handling to Cameras Collation Function

[MingwuZheng]

This pull request addresses an issue in the nerfstudio_collate function where the metadata of Cameras objects is not properly accounted for during batch processing.

In the existing implementation, the metadata of Cameras objects is not included during collation, leading to the assignment of None to the metadata field. This may cause downstream issues for components that rely on this metadata for processing.

The proposed changes aim to ensure that the metadata of each Cameras object in a batch is properly collated. The updates involve the creation of a metadata dictionary that collates the metadata from each Cameras object in the batch. This dictionary is then passed to the Cameras constructor during batch assembly.

The changes ensure that all Cameras objects have the same metadata keys before proceeding with the collation. This ensures consistency across the batch and prevents potential errors from occurring due to mismatched or missing metadata keys.

[tancik]

Can you provide a PR description. Also look here for installing pre-commit to avoid these failing checks.

[MingwuZheng]

Can you provide a PR description. Also look here for installing pre-commit to avoid these failing checks.

My apologies for the initial oversight. I have now updated the PR description to provide more clarity on the changes being introduced.

Additionally, I would like to point out potential bugs in the current code. In the following section:

distortion_params=op(
    [
        cameras.distortion_params
        if cameras.distortion_params is not None
        else torch.zeros_like(cameras.distortion_params)
        for cameras in batch
    ],
    dim=0,
),
camera_type=op([cameras.camera_type for cameras in batch], dim=0),
times=torch.stack(
    [cameras.times if cameras.times is not None else -torch.ones_like(cameras.times) for cameras in batch],
    dim=0,
),

If cameras.distortion_params or cameras.times are None, the else clause attempts to execute torch.zeros_like(cameras.distortion_params) or -torch.ones_like(cameras.times) respectively. However, zeros_like and ones_like do not support None as an input. This could lead to an error during execution.

Regarding the failing checks, I'll ensure to install and run pre-commit as suggested in the contributing guidelines. Thank you for bringing this to my attention.

[MingwuZheng]

Hey @tancik,
Just wanted to let you know that I've passed the tests and updated the PR. Could you take another look when you have a moment?

[tancik]

LGTM