Submit a dataset

A rig we do not have is more useful than another run on a rig we do. Datasets are what make the benchmark meaningful, and contributed ones are cited and credited.

1. Check it is usable

A dataset only helps if a calibration can actually be run on it:

• Raw, synchronised sensor streams — no pre-rectified images, no vendor-corrected IMU output. Corrections you cannot undo make intrinsics uncalibratable.
• A visible calibration target for a meaningful part of the recording, with its geometry stated (tag family, tag size, spacing).
• Well-excited motion — rotation about all three axes, not just translation. See Degeneracy.
• A static period at the start so IMU initialisation has something to work with.
• A licence you have the right to grant.

2. Package it

Item	Requirement
Data	ROS bag or EuroC layout
Target	Tag family, tag size, spacing, board dimensions
Sensors	Model, rate, resolution, and topic for each
Timing	Whether streams are hardware-synchronised; rolling-shutter convention if applicable
Reference	Any existing calibration, and what produced it

If a reference calibration exists — even an imperfect one — include it and say how it was obtained. It turns the dataset into something a method can be validated against rather than merely run on.

3. Write a dataset card

One page, alongside the data:

• What the rig is and what makes it interesting
• How and where the recording was made
• Known problems — dropped frames, clock glitches, saturated segments, soft-synchronised streams
• Licence and how to cite it
• Contact for questions

Say what is wrong with it

The known-problems section is the most valuable part. A documented flaw is a usable dataset; an undocumented one produces results nobody can explain, and costs every later user the same debugging.

4. Licensing

Datasets need a licence that permits redistribution and derivative results — CC BY 4.0 is the usual choice. Recordings containing identifiable people or private spaces need that resolved before submission, not after.

5. Send it

Open an issue on the site repository describing the dataset and linking the data. Hosting is arranged from there: large datasets go to Hugging Face with a Zenodo DOI so they are citable.

Credit

Contributed datasets get a per-dataset DOI and citation, a named entry on Contributors, and — where the dataset drives a benchmark or dataset paper — co-authorship. Details in Governance.

To submit a result on an existing dataset instead, see Submit results.