In EuRoC V101 dataset, the estimated trajectory has a big error!

[highlightz]

Hi, I use the default pinhole.yaml config file and test V101 dataset with my machine Intel® Core™ i5-4570 CPU @ 3.20GHz × 4 . Sometimes the estimator runs through to the end without tracking failed, sometimes, it cannot run to the end. For the first situation, I've got the below trajectory:

And for the latter situation, I got the below trajectory, with a big error:

scale: 2.650539
compared_pose_pairs 1651 pairs
absolute_translational_error.rmse 0.539003 m
absolute_translational_error.mean 0.464468 m
absolute_translational_error.median 0.451905 m
absolute_translational_error.std 0.273484 m
absolute_translational_error.min 0.101481 m
absolute_translational_error.max 1.536845 m
How can I improve the performance by tuning the parameters?
As a matter of fact, I notice in your paper that, your SVO2 program runs well on all EuRoC datasets. Shown as below chart:

[zhangzichao]

Hi, you will need to change the parameters. I will add some example parameter files for EUROC as well as a general instruction for parameters soon.

[highlightz]

Okay, Thanks.

[zhangzichao]

Hi, we added some example launch files and parameters for EuRoC datasets. Should be better than the default one and also helps with the issue #3 .

[JohanVer]

Hi, I got everything to run and just tried the EuRoC dataset MH_01_easy in stereo mode using the launch file: euroc_stereo_imu.launch .
Thereby I got some questions:
1.Though it tracks the position of the camera I get lots of: "seed sigma is nan!-0.338734, sq-nan, check-convergence = 0" and "critical". What does that mean?
2. I also wondered if the baseline of the stereo isn't provided somewhere (does it estimate scale in stereo mode)?
3. Is there there any need to pre-undistort or rectify the input images or is it done by the tool?

Thanks a lot for the great software!

[zhangzichao]

Hi,

1. it means something wrong with the depth filter, we know this issue and hope to fix/suppress it in the future.
2. the baseline is defined implicitly in the calibration file. The T_B_C defines the transformation from the body frame to the camera frame (transformation convention).
3. No, there is no need for undistortion. The camera distortion is already taken care of internally.

[JohanVer]

Hey, great thank you very much for your fast answer!

1. Can I for now ignore the error or is it known that this problem influences the tracking accuracy significantly?
2. Regarding the baseline:
   I have kalibr calibration files (yaml) and converted them to the svo format using your provided script. The kalibr calibration includes the matrix T_cn_cnm1 which should give the transformation between the right camera and the left camera. The converter sets both T_B_C matrices to identity because of "No IMU is specified, so T_B_C is set to identity". So I should overwrite the T_B_C of the right camera in the svo-yaml with T_cn_cnm1 and leave the left T_B_C with the identity matrix? Is this correct?