Publish the camera calibration material as a separate dataset
#2
by retttx - opened
Proposal
Publish Perceive's camera calibration material as a separate dataset, so that this one stays focused on recordings of people in stores and the calibration data is there for anyone who wants to test calibration algorithms rather than run the tracker.
Why separate
The stereo heads Perceive built are unique hardware, so the calibration files are not needed to use the tracking data: everything the pipeline needs for the published scenarios is already bundled in calibration/ here and in the multiview repository. What the archive holds beyond that is the raw material the calibrations were derived from, which is only interesting to people working on calibration itself.
What exists (from the archive inventory)
- Distortion calibration photos per camera and sensor: about 9,000 JPG/PNG stills of the ArUco cube and calibration boards, plus a few long calibration videos (about 43 GB).
- Calibration run outputs: per-camera stereo rectification results, reprojection images and scene fits (about 47 GB of PNG, mostly diagnostic renders; the useful subset is much smaller).
- The compact outputs already here: distortion models (polynomial, not the OpenCV radial model), stereo extrinsics, ArUco cube definitions, scene camera positions (under 1 GB).
Open questions
- Scope: just the stills and the ground truth needed to re-run the distortion and stereo calibration, or also the diagnostic renders?
- Format and size target: raw stills are large; a per-camera subset (say 30 to 50 images) would cover the algorithms without tens of GB.
- License: same CC BY-NC 4.0 or something more permissive, since there are no people in calibration imagery (to be verified image by image).
- Documentation: the calibration procedure is described in the
multiviewrepository docs (src/perceive/dox/); the dataset card should link to it and give one worked example per stage.
Suggested name
prcvlabs/multiview-calibration