search

claydergc / loam_velodyne_kitti_ros

LOAM algorithm to use with the KITTI Dataset
47 stars 12 forks source link

the accuracy of pose in kitti training data #10

Open cgbcgb opened 4 years ago

cgbcgb commented 4 years ago

the translation error can reach 0.88% as described in Zhang Ji ’s paper？

claydergc commented 4 years ago

I haven't still tried the metric that the KITTI website uses. I'm still working on determining which parameters of the code affect the results the most. These days, I've been modifying some constants of the LM optimization which seem to affect the results. What I have up to here is the following figure of sequence 00.

sequence00

cgbcgb commented 4 years ago

your result of sequence 00 is pretty good ！！！ I download the evaluation code from the kitti official website to quantitatively evaluate the accuracy calculated by A-LOAM. And recently I changed the way to calculate scanID and design weight for the residual considering the number of edge、plane features. Thus I reduced the translation error of sequence 00 from 0.83% to 0.76%, but still worse than your result from your picture, my result is as follows: 00 Can you share what parameters you have changed in the LM optimization? By the way，the code of LOAM used Gauss Newton method instead of LM method, right ?

claydergc commented 4 years ago

The parameters I used are in this repository. However, they may differ depending on the speed of your computer. The first thing you have to determine is the correct rate to publish point clouds of KITTI sequences. This is important because sometimes the code can skip some point clouds while running.

I am interested in changes you've done related with the scan ID and residuals. Maybe, we could share a GitHub repo and merge our changes to see if finally we can reproduce completely J. Zhang's results.

Let me know if you're interested on my proposal.

cgbcgb commented 4 years ago

The parameters I used are in this repository. However, they may differ depending on the speed of your computer. The first thing you have to determine is the correct rate to publish point clouds of KITTI sequences. This is important because sometimes the code can skip some point clouds while running.

I am interested in changes you've done related with the scan ID and residuals. Maybe, we could share a GitHub repo and merge our changes to see if finally we can reproduce completely J. Zhang's results.

Let me know if you're interested on my proposal.

OK, I will send to you an email later to communicate in detail