metadata

license: cc-by-4.0

Monado SLAM Datasets

The Monado SLAM datasets (MSD), are egocentric visual-inertial SLAM datasets recorded for improving the Basalt-based inside-out tracking component of the Monado project. These have a permissive license CC-BY 4.0, meaning you can use them for any purpose you want, including commercial, and only a mention of the original project is required. The creation of these datasets was supported by Collabora

Monado is an open-source OpenXR runtime that you can use to make devices OpenXR compatible. It also provides drivers for different existing hardware thanks to different contributors in the community creating drivers for it. Monado provides different XR-related modules that these drivers can use. To be more specific, inside-out head tracking is one of those modules and, while you can use different tracking systems, the main system is a fork of Basalt. Creating a good open-source tracking solution requires a solid measurement pipeline to understand how changes in the system affect tracking quality. For this reason, the creation of these datasets was essential.

These datasets are very specific to the XR use-case as they contain VI-SLAM footage recorded from devices such as VR headsets but other devices like phones or AR glasses might be added in the future. These were made since current SLAM datasets like EuRoC or TUM-VI were not specific enough for XR, or they didn't have permissively enough usage licenses.

For questions or comments you can use the Hugging Face Community, join Monado's discord server and ask in the #slam channel, or send an email to [email protected].

List of sequences

https://huggingface.co/datasets/collabora/monado-slam-datasets/resolve/main/M_monado_datasets/MI_valve_index/extras/previews/MIC01_camcalib1.webm

MI_valve_index

Valve Index datasets

These datasets were recorded using a Valve Index with the vive driver in Monado and they have groundtruth from 3 lighthouses tracking the headset through the proprietary OpenVR implementation provided by SteamVR. The exact commit used in Monado at the time of recording is a4e7765d. The datasets are in the ASL dataset format, the same as the EuRoC datasets. Besides the main EuRoC format files we provide some extra files with raw timestamp data for exploring realtime timestamp alignment techniques.

The dataset is postprocessed to reduce as much as possible special treatment from SLAM systems: camera-IMU and groundtruth-IMU timestamp alignment, IMU alignment and bias calibration has been applied, lighthouse tracked pose has been converted to IMU pose and so on. Most of the post processing was done with Basalt calibration and alignment tools, as well as the xrtslam-metrics scripts for Monado tracking. The postprocessing process is documented in this video which goes through making the MIPB08 dataset ready for use starting from its raw version.

Data

Camera samples

In the vive driver from Monado we don't have direct access to the camera device timestamps but only to V4L2 timestamps. These are not exactly hardware timestamps and have some offset with respect to the device clock in which the IMU samples are timestamped.

The camera frames can be found in the camX/data directory as PNG files with names corresponding to the their V4L2 timestamps. The camX/data.csv file contains aligned timestamp of each frame. The camX/data.extra.csv also contains the original V4L2 timestamp and the "host timestamp" which is the time at which the host computer had the frame ready to use after USB transmission. By separating arrival time and exposure time algorithms can be made to be more robust for real-time operation.

The cameras of the Valve Index are global shutter with a resolution of 960x960 streaming at 54fps. They have autoexposure enabled. While the cameras of the Index are RGB you will find only grayscale images in these datasets. The original images are provided in YUYV422 format but only the luma component is stored.

For each dataset, the camera timestamps are aligned with respect to IMU timestamps by running visual-only odometry with Basalt on a 30-second subset of the dataset. The resulting trajectory is then aligned with the basalt_time_alignment tool that aligns the rotational velocities of the trajectory with the gyroscope samples and returns the resulting offset in nanoseconds. That correction is then applied to the dataset. Refer to the postprocessing walkthrough video for more details.

IMU samples

The IMU timestamps are device timestamps, they come at about 1000Hz. We provide an imu0/data.raw.csv file that contains the raw measurements without any axis scale-misalignment nor bias correction. imu0/data.csv has the scale-misalignment and bias corrections applied so that the SLAM system can ignore those corrections. imu0/data.extra.csv contains the arrival time of the IMU sample to the host computer for algorithms that want to adapt themselves to work on real-time.

Groundtruth information

The groundtruth setup consists of three lighthouses 2.0 base stations and a SteamVR session providing tracking data through the OpenVR API to Monado. While not as precise as a other MoCap tracking systems like OptiTrack or Vicon it should still provide pretty good accuracy and precision close to the 1mm range. There are different attempts at studying the accuracy of SteamVR tracking that you can checkout like this, this, or this. When a tracking system gets closer to milimiter accuracy these datasets will no longer be as useful for improving it.

The raw groundtruth data is stored in gt/data.raw.csv. OpenVR does not provide timestamps and as such, the timestamps recorded are from when the host asks OpenVR for the latest pose with a call to GetDeviceToAbsoluteTrackingPose. The poses contained in this file are not of the IMU but of the headset origin as interpreted by SteamVR, which usually is between the middle of the eyes and facing towards the displays. The file gt/data.csv corrects each entry of the previous file with timestamps aligned with the IMU clock and poses of the IMU instead of this headset origin.

Calibration

There are multiple calibration datasets in the MIC_calibration directory. There are camera-focused and IMU-focused calibration datasets. See the README.md there for more information on what each sequence is.

In the MI_valve_index/extras directory you can find the following files:

calibration.json: Calibration file produced with the basalt_calibrate_imu tool from MIC01_camcalib1 and MIC04_imucalib1 datasets with camera-IMU time offset and IMU bias/misalignment info removed so that it works with the fully the all the datasets by default which are fully postprocessed and don't require those fields.
calibration.extra.json: Same as calibration.json but with the cam-IMU time offset and IMU bias and misalignment information filled in.
factory.json: JSON exposed by the headset used for recording with information from factory that include calibration and other data. It's not used for anything but might be of interest.
other_calibrations/: Results from calibrating using the other datasets for comparisson and checking most of them are similar. MICXX_camcalibY have camera only calibration produced with the basalt_calibrate tool, while the corresponding MICXX_imucalibY datasets use these datasets as a starting point and have the basalt_calibrate_imu calibration results.

Camera model

By default, the calibration.json file provides parameters k1, k2, k3, and k4 for the Kannala-Brandt camera model with fish-eye distortion (also known as OpenCV's fish-eye).

Calibrations with other camera models might be added later on, otherwise you can use the calibration sequences for custom calibrations.

IMU model

For the default calibration.json where all parameters are zero you can ignore any model and just use the measurements present in imu0/data.csv directly. If instead you want to use the raw measurements from imu0/data.raw.csv you will need to apply the Basalt accelerometer and gyroscope models that uses a misalignment-scale correction matrix together with a constant initial bias. The random walk and white noise parameters were not computed and default reasonable values are used instead.

Post-processing walkthrough

If you are interested in understanding the step-by-step procedure of postprocessing of the dataset, below is a video detailing the procedure for the MIPB08 dataset.

Sequences

MIC_calibration: Calibration sequences recording this calibration target from Kalibr with the squares of the target having sides of 3cm. Some sequeneces are focused on camera calibration covering the image planes of both stereo cameras while others on IMU calibration properly exciting all six components of the IMU.
MIP_playing: Datasets in which the user is playing a particular VR game on SteamVR while Monado records the datasets.
- MIPB_beat_saber: This contains different songs played at different speeds. The fitbeat song is one that requires a lot of head movement while MIPB08 is a long 40min dataset with many levels played.
- MIPP_pistol_whip: This is a shooting and music game, each dataset is a different level/song.
- MIPT_thrill_of_the_fight: This is a boxing game.
MIO_others: These are other datasets that might be useful, they include playpretend scenarios in which the user supposed to be playing some particular game, then there is some inspection and scanning/mapping of the room, some very short and lightweight datasets for quick testing, and some datasets with a lot of movement around the environment.

License

This work is licensed under a Creative Commons Attribution 4.0 International License.