Alignment and rescale of trajectories in EuRoC dataset

[squaresoft2015]

The generated trajectories have different number of poses with ground truth. For example, ORB-SLAM3 generates 2674 poses in monocular mode and 2700 poses in stereo mode. But the officially provided dataset contains 26302 poses in ground truth file. How can I align the estimated poses with ground truth? I tried -va parameter but it promots: [ERROR] found no matching timestamps between reference and f_dataset-V101_mono.txt with max. time diff 0.01 (s) and time offset 0.0 (s) Any help? Thanks

[richard-elvira]

ORBSLAM3 only reports the trajectory at frame level, that means we only obtain the pose when an image exists. In the other hand, EuRoc provides a GT with other measure devices, in Machine Hall uses Leica and in Vicon Room an OptiTrack. For that reason the amount of poses is not the same.

[squaresoft2015]

ORBSLAM3 only reports the trajectory at frame level, that means we only obtain the pose when an image exists. In the other hand, EuRoc provides a GT with other measure devices, in Machine Hall uses Leica and in Vicon Room an OptiTrack. For that reason the amount of poses is not the same.

Thank you so much! I have converted the groundtruh to TUM format. But when I plot the two trajectories in the same coordinate with rescale and alignment option, the evo program promopt: [ERROR] found no matching timestamps between reference and f_dataset-MH01_mono.txt with max. time diff 0.01 (s) and time offset 0.0 (s) How can I align them since the time stamp may not be in the corresponding position?

[richard-elvira]

Check the timestamp units, perhaps one of the files has the timestamps in seconds and the other in nanoseconds.

[squaresoft2015]

Check the timestamp units, perhaps one of the files has the timestamps in seconds and the other in nanoseconds.

Yes, the generated trajectories have time stamp like this: 1403636579863555584.000000 0.000043671 0.008587432 0.003351578 -0.000423142 0.004911167 -0.002808022 0.999983907 But the converted ground truth has time stamp like this: 1.403637132888319016e+09 4.631116999999999706e+00 -1.786812000000000067e+00 5.771129999999999871e-01 -1.563670000000000060e-01 -7.763449999999999518e-01 -7.222900000000000154e-02 6.063169999999999948e-01 I converted the first entry into the format of "1403636580838555.000000" and the TUM style ground truth is: 1403636580838555.000000 4.688319 -1.786938 0.783338 -0.082152 -0.153029 -0.827383 0.534108 The evaluation tool still prompts the same error.

[richard-elvira]

The timestamp units are differents. 1403636579863555584.000000 // Timestamp of the trajectory 1403636580838555.000000 // Timestamp in your converted GT

You need to multiply the timestamps in your converted GT by 1e3 to obtain the same unit than the trajectory timestamps.

[squaresoft2015]

The timestamp units are differents. 1403636579863555584.000000 // Timestamp of the trajectory 1403636580838555.000000 // Timestamp in your converted GT

You need to multiply the timestamps in your converted GT by 1e3 to obtain the same unit than the trajectory timestamps.

Thank you for your reply! I have multiplied the timastamps by 1000, but the error remains the same. I notice the timestamp in data.csv (ground truth) is 1403637132888310000 and the corresponding timestamp in trajectory file is 1403636579863555584.000000. The two digits vary a lot.

[254843]

Hey! I also met the problem of real pose and estimated pose timestamp. Have you solved it? @squaresoft2015

[dinarkino]

Hey! I also met the problem of real pose and estimated pose timestamp. Have you solved it? @squaresoft2015

Hey, I solved the problem by using an advice from here. You should do something like that with an estimated pose with the help of evo package or manually.

from evo.tools import file_interface

traj = file_interface.read_tum_trajectory_file("vins_V102.txt")

traj.timestamps = traj.timestamps / 1e9

file_interface.write_tum_trajectory_file("vins_V102_clean.txt", traj)

[jdtardos]

Pleas, check version v1.0