Question about the fitness change about fast_gicp HOT 5 CLOSED

This is because the old one had some bug and showed wrong fitness scores. I confirmed that the registration accuracy on KITTI is not changed by the optimizer update.

from fast_gicp.

Thanks, I am glad it was fixed.

from fast_gicp.

This is because the old one had some bug and showed wrong fitness scores. I confirmed that the registration accuracy on KITTI is not changed by the optimizer update.

Hi, @koide3

I want to know more about the fitness score of the registration. I've read the codes behind it and know it estimates the average dist between aligned and target cloud. But it is weird that the score may be large though the result is good. So I wonder how exactly to measure the goodness of registration? For example if I need to filter out the bad result. Thanks.

Looking forward to your advice on that ;)

from fast_gicp.

Actually, the fitness_score (RMSE between points) is not really a good metric to evaluate registration results, and this is why some extensions of ICP (e.g., normal ICP, GICP), which minimize the distance w.r.t. the object surface, perform better than the pure ICP. IMO, a reasonable way to assess the accuracy of registration results is evaluating the accuracy of estimated trajectories with the registration methods.

BTW, if you just need to find largely corrupted registration results, overlap rate can be used.

from fast_gicp.

Thank you, @koide3

Actually i want to incorporate it for loop closure verificaition. Overlap is also what i can came out.

from fast_gicp.