Large Data Alignment Error Using The Valve Index
Hey Mateo,
We've spoken via email previously but I am re-iterating the problem for the discussion forum.
I am using Basalt and the Valve Index as a part of a university project comparing the tracking paths between Lighthouse tracking and the stereo camera set up on the headset itself, and I'm running into a couple of issues when comparing.
I was unable to capture data using Monado and as a result, I have created my own python script to generate the required data in the EuRoC format. As a result, I have captured the lighthouse positional data to use as ground truth, the IMU sensor readings from the Index itself and images from both cameras using the same timestamp.
When running the data through basalt, the comparison between GT and the generated tracking produce a large error and are not as similar as your recorded datasets. When I run your sets calibration file it does slightly improve.
I was just wondering where I went wrong in the process and if there was anything I could do to fix this and record my own data, with the precision of yours, within the next couple of weeks.
I was unable to upload a dataset as a zip file so I have stored it in a Google Drive folder for now. Once downloaded please let me know so I can remove it/ if you require more data please let me know also.
-Link Removed-
In addition, I have provided two images showing the issue (one image from the dataset you've created and one from mine)
Kind Regards,
NKoops
Hello, so after checking the datasets there are a couple of things that look off:
- The camera calibration extrinsics are definitely off, the distance between the two cameras is too big compared to a working calibration.
- The two calibration files you provide have a different
imu_update_rate, one of ~400 and the other of ~600, and when checking witheuroc/euroc_ops.py get_max_sensor_dt imu DATASETfrom xrtslam-metrics it seems your dataset IMU samples are not being consistent compared to the Monado SLAM datasets which pretty consistently save IMU samples every 1ms (1000000ns), even on the longest 40min dataset.
~/D/p/d/i/n/DataTesting (main|✚7…) $ $xrtmet/euroc/euroc_ops.py get_max_sensor_dt imu StraightLine/
Warning: gt path doesn't exist
Max sensor deltatime=15055872 on line 5344 and next
Min sensor deltatime=1641216 on line 7330 and next
~/D/p/d/i/n/DataTesting (main|✚7…) $ $xrtmet/euroc/euroc_ops.py get_max_sensor_dt imu 3mBy3m
Warning: gt path doesn't exist
Max sensor deltatime=12513024 on line 802 and next
Min sensor deltatime=1181440 on line 2094 and next
~/D/p/d/i/n/DataTesting (main|✚7…) $ $xrtmet/euroc/euroc_ops.py get_max_sensor_dt imu $msdmi/MIP_playing/MIPB_beat_saber/MIPB08_beatsaber_long_session_1
Max sensor deltatime=1000020 on line 1882523 and next
Min sensor deltatime=999895 on line 2 and next
I would recommend you two things, check how you are saving the IMU samples, it seems off and Basalt expects samples that come at a more or less consistent rate, this is probably making your IMU-camera calibration off. Also, check that your camera-only calibration json file is good by comparing it with the ones provided in MSD (the ones named camcalib*), check the px distance for example, it should be around 13 cm. Finally, check if you are applying some artificial transformation to the IMU samples of the Index.
Maybe trying to make Monado work for recording the EuRoC dataset with OpenVR tracking data through SteamVR (or libsurvive) as we did for the Monado SLAM datasets might be a good idea. This is the branch I used which doesn't diverge too much from current Monado albeit it does try to remove any artificial modification of timestamps before saving them to disk: https://gitlab.freedesktop.org/mateosss/monado/-/tree/mateosss/idxdataset?ref_type=heads.