search

MIT-SPARK / TEASER-plusplus

A fast and robust point cloud registration library
MIT License
1.69k stars 333 forks source link

[QUESTION]Inconsistent results between c++ and python implementation #141

Closed stephen9999999 closed 2 months ago

stephen9999999 commented 1 year ago

Have you read the documentation?

Post your theoretical questions / usage questions here.

Hi, thanks for your great work. I have test teaser_cpp_fpfh and teaser_python_fpfh_icp with the same pcd files, but the results are inconsistent. Is there any difference of this two implementation?

For teaser_python_fpfh_icp, I test the algorithm with two pcd files collected by myself. And I get the reasonable result. The estimated transforms are: T_teaser:
[[ 0.93723273 0.11378536 -0.3296175 -0.09877629] [-0.10340515 0.99343573 0.04891655 0.02664667] [ 0.33301979 -0.01176205 0.94284647 0.10280838] [ 0. 0. 0. 1. ]]

T_icp_refine: [[ 0.93582733 -0.00977672 -0.35232319 0.10150537] [ 0.02057001 0.99942636 0.02690389 0.02983166] [ 0.35185805 -0.03242469 0.93549161 0.08510608] [ 0. 0. 0. 1. ]]

For teaser_cpp_fpfh, I slightly modify the code to input two pcd files. And the estimated transform is: Estimated rotation: 0.341367 -0.376026 0.861437 0.451216 0.869541 0.200758 -0.824544 0.320162 0.466501 Estimated translation: -4.24419 -0.724059 2.3231 The estimated transform is not reasonable.

May I ask why the result of c++ implementation is not reasonable? how can I improve the result of c++ version?

jingnanshi commented 1 year ago

Make sure in the C++ implementation to call reset after every solve using the solver.

stephen9999999 commented 1 year ago

Thanks for your reply. In the demo teaser_cpp_fpfh.cc, the solver is called only once. I input tow pcd files, and use the solver to calculate the transform of the two point cloud. I follow your suggestion and call reset after every solve

2022-08-20 09-53-44 的屏幕截图

but I get an error: "error: no matching function for call to ‘teaser::RobustRegistrationSolver::reset()’". what should I do to improve the result of c++ implementation? looking forward to your suggestion.

jingnanshi commented 1 year ago

See this line: https://github.com/MIT-SPARK/TEASER-plusplus/blob/111e2ce81aa8c25fad66fcd1feecb758e1c76281/teaser/include/teaser/registration.h#L738

Can you also try using solve(src_correspondences, tgt_correspondences) instead of passing the correspondences in separately? Basically just make sure the columns of src_correspondences and tgt_correspondences are corresponding 3D points.

LimHyungTae commented 2 months ago

If there are no further questions, we will close the issue :) Thx