pablovela5620/msd-rrd

Monado SLAM Dataset as Rerun recordings

Every recording of the Monado SLAM Dataset (Valve Index, HP Reverb G2, Samsung Odyssey+; 64 recordings, 316 minutes) converted to Rerun .rrd files by the dataforge package of rerun-io/examples-monorepo. Made for slam-rs, which replays these files from a local Rerun catalog.

Layout

Three layers per recording, sharing one recording id, plus the viewer blueprints:

base/<recording_id>.rrd             per-camera AV1 video, gyro and accel at 1 kHz, magnetometer (G2, Odyssey+), rig calibration
gt/<recording_id>.rrd               Lighthouse ground-truth poses at 1 kHz on the rig node, trajectory line
sensor_metadata/<recording_id>.rrd  IMU noise densities, bias random walks, applied time shift
blueprints/msd-<device>.rbl         default viewer layout per headset; msd-<device>-table.rbl for the table view

<recording_id> is msd-<device>__<collection>__<sequence>, for example msd-index__MIO_others__MIO10_short_2_panorama. Devices: index (33 recordings, 2 cameras), g2 (15, 4 cameras), odyssey (16, 2 cameras). Entity paths follow the exoego:v2 rig schema (/world/rig_00/cam_NN/pinhole/video, /world/rig_00/imu_00/{gyro,accel}); the timeline is video_time in nanoseconds since the capture start.

Use

Register the three directories as datasets msd-index, msd-g2, msd-odyssey on a local rerun server; the slam-rs README shows the commands. A single recording can also be replayed straight from the file pair (base + gt) without a catalog.

Conversion: the upstream PNG frames were encoded to AV1 upright per the factory calibration, IMU and ground truth were kept at native rate, and the ground-truth world-up axis was measured rather than assumed. Details in packages/dataforge/docs/msd.md of the monorepo.