Skip to content

Question: Are per-timestamp camera extrinsics intentionally omitted in the public ABC-130k release? #7

Description

@labs-sunzhengmao

Hi ABC team,

First of all, thank you for open-sourcing the dataset and training pipeline.

I have a question about camera calibration in the public release.
From the released conversion and training code, it looks like training uses:

states_actions.bin
combined_camera-images-rgb.mp4
episode_metadata.json (task name, camera list, timing info, etc.)
I could not find per-timestamp camera extrinsics being exported or loaded during training.

Could you clarify whether this is intentional for the current release?
Specifically:

Are per-timestamp camera extrinsics available anywhere in the raw data (for example in MCAP topics/metadata) but just not included in the converted training format?
If not currently available, are there plans to release them in a future update?
If the camera rig is assumed static, is there a recommended canonical extrinsic for each camera that users should use for geometry-aware methods?
This would be very helpful for users working on multi-view geometric reasoning, 3D reconstruction, or camera-aware policy learning.

Thanks again for the great release.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions