Question About Loop Closure in /vins_estimator/camera_pose Results

[setareh-soltanieh]

Hi,

I have been working on this algorithm as part of my Master's thesis and have tested it using the EuRoC dataset with the provided configurations from this repository. I ran each test case 10 times, recorded the final poses from the /vins_estimator/camera_pose topic, and computed the average Absolute Trajectory Error (ATE) for each test case.

However, I noticed that my results are closer to the VINS results provided in the paper without loop closure, even though I ensured that loop_closure: 1 was set in the configuration file.

In RViz, I can see loop closure happening (e.g., matched frames and camera pose jumps), which suggests that loop closures are being detected. Despite this, the results from /vins_estimator/camera_pose do not seem to reflect the improvements expected from loop closure.

Can someone clarify why the poses from the /vins_estimator/camera_pose topic appear to lack the effects of loop closure, or if there is another topic I should be using to evaluate the loop-closed trajectory?

Thank you for your help!