Giters

jedeschaud / ct_icp

CT-ICP: Continuous-Time LiDAR Odometry

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

why my Elapsed Search Neighbors cost too much? can you answer for me,thanks

dongfangzhou1108 opened this issue · comments

commented

thanks for openning your code.

Number of points in sub-sampled frame: 35679 / 126902
Initial ego-motion distance: 0
Elapsed Normals: 35.0737
Elapsed Search Neighbors: 567.692
Elapsed A Construction: 0
Elapsed Select closest: 2.18221
Elapsed Solve: 0.222377
Elapsed Solve: 8.04642
Number iterations CT-ICP : 5
Elapsed Elastic_ICP: 681.094
Number of Keypoints extracted: 4640 / Actual number of residuals: 4180
Trajectory correction [begin(t) - end(t-1)]: 0
Final ego-motion distance: 0.29746
Average Load Factor (Map): 0.324829
Number of Buckets (Map): 16384
Number of points (Map): 39514
Elapsed Time: 683.08 (ms)

here is my log info of one frame, why my Elapsed Search Neighbor cost too much?

best wishes

commented

and i play by default config file

commented

i guess first 20 frames cost much?

commented

here is my kitti raw 09 output
image
can see it lost a lot...

commented

i modify your code on ros,have a bad output, but when on your own code , it is no problem,

commented

Hi, thank you for your interest

For your first question, for the first 20 frames (the time to construct the map), we use a larger neighborhood (inspecting the 5^3=125 closest voxels instead of the 3^3=27 neighbors). This requires more time for the first 20 frames, but insure a good initialization.

I do not really understand your second question, but we will work on our own ROS support in a month or so.
Your results on kitti_raw sequence 09 seem very inconsistent with what we observe on the dataset running our code.
You can use our dataset of PLY files (link in the Readme), and you should have much cleaner results.

Best regards,

commented

Hi, thank you for your interest

For your first question, for the first 20 frames (the time to construct the map), we use a larger neighborhood (inspecting the 5^3=125 closest voxels instead of the 3^3=27 neighbors). This requires more time for the first 20 frames, but insure a good initialization.

I do not really understand your second question, but we will work on our own ROS support in a month or so. Your results on kitti_raw sequence 09 seem very inconsistent with what we observe on the dataset running our code. You can use our dataset of PLY files (link in the Readme), and you should have much cleaner results.

Best regards,

sorry, i do not make time from 0 to 1, when i change this, everything goes right.
thanks for your reply.

best wishes