To save storage space, the resolution of the ZED2i stereo images has been reduced to 640×360 in resolution since Dec 8 2023.
In some sequences, the ZED2i images and IMU data have lower rates than the nominal values, likely due to low power voltage.
Residual time corrections: The residual time corrections are given at here with an accompanying readme copied below:
residual_td.csv contains the residual time offset corrections due to several bugs.
These bugs include the KISS-ICP timestamp, and the central Gaussian smoother.
Due to the KISS-ICP bug, the Hesai PandarXT-32 Point cloud timestamps should decrease by 0.05 sec to align with the other topics, e.g., X36D IMU.
This is due to the fact that the KISS-ICP outputs motion at t + dt / 2 but stamps it with t and dt = 100 ms for PandarXT-32.
For sequence 20230920/data2, the point cloud timestamps should instead increase by 0.05 sec as its lidar time offset -0.07932 sec was identified as an outlier.
Oddly, for this sequence, the correlation algorithm result shows no problem by visual check.
Just like time_corrections_set1 and set2.csv, these residual td have been applied to the sequence data, including both topics in rosbags and files in data folders.
Because the reference trajectories and the full trajectories are generated using the lidar timestamps before applying the residual td, for accurate benchmarking, we need to correct the reference trajectories and full trajectories timestamps by plus the residual td for /hesai/pandar given in residual_td.csv. We are working on this, and will release the corrected trajectories in a few days.
However, by comparing the time offsets estimated by using ORB-SLAM3 mono + zed2i IMU, and swift_vio for some sequences covering all three platforms, I found that their time offset estimates are pretty close. The comparison table is at here.