Similar to `reg.getFitnessScore();

[bear-rabbit]

Hello! We have used your previous work, Fast_GICP, in our project and it performed very well. Therefore, we would like to try your new work, small_gicp. However, does it have a function similar to reg.getFitnessScore();?

[koide3]

In RegistrationResult, there are two metrics that may be used as an alternative for getFitnessScore().

• num_inliers represents the number of source points with valid correspondences (i.e., "num of overlapping points").
• error represents the final registration error (sum of distances between corresponding points)

[bear-rabbit]

Hello! I’m glad that the metrics you provided are usable. However, we still have one question: both smallgicp and fastgicp exhibit a significant fluctuation in runtime, varying from approximately 100ms to 1000ms on the same device running the same program（Average Times）. I’m unsure how to adjust the program to achieve maximum efficiency.

[koide3]

It may stem from data association, which sometimes causes an oscillation of the objective function. A quick fix is to set the maximum optimization iteration to a small number (e.g., 10 is sufficient for most cases). If you could provide example data, I can try implementing an oscillation compensation algorithm.

[bear-rabbit]

Okay, I tried the method you suggested (reducing the number of iterations), and it indeed achieved a relatively good result. In fact, we are using this algorithm as the final step in our relocalization algorithm. The effect might not be very noticeable when applied to a single dataset, but the change in overall processing time can be quite significant. The following two point cloud frames are from our test data: on our i7 10th generation device, the processing time is approximately 400ms to 900ms (with 10 iterations).
data.zip