Inquiry Regarding Dataset Processing Methodology

[setareh-soltanieh]

Hello,

I'm impressed with the excellent work you've done. I'd like to test your network on a different dataset and was wondering if you could provide some insights into the process of calculating 'calibration.json', 'imu0_resampled_description.json', and 'imu0_resampled.npy' from the 'imu_samples_0.csv' file. I understand that your raw data consists solely of IMU measurements collected from a mobile device. Could you shed some light on the methodology behind generating these specific files?

Thank you!

[losehu]

Have you solved this problem? I have the same question.

[setareh-soltanieh]

Hi @losehu
I reached out to one of the contributors, but unfortunately, they no longer have the code for that component—which is essential if you plan to use this network. So to answer your question: no, I haven’t been able to resolve this issue.
As an alternative, I recommend looking into other networks like OpenVINS, which provide complete preprocessing pipelines and support real-time data.

[Aljen-511]

After reading the closed issues and the calibration.json files within the newest published dataset, I think the calibration.json records the calibrated IMU states (e.g. , accelerometer bias and so on), and that calibrated information is used for training. While the imu_samples_0.csv files contain the uncalibrated raw imu data, which is the input of the kalman filter framework.