[QUESTION] "Edge exists" Error when Processing Point Cloud with TEASER Solver

[WuKyrin]

I encountered an "Edge exists" error message while running the TEASER solver on a point cloud dataset in my project. This issue arises under specific circumstances, which leads me to believe there might be a problem with how edges are being processed or identified within the dataset. This dataset is a stl file of femur and the ply file we captured with our RGBD camera of the operating scences of patients.

[arminkmbr]

I get the same error followed by a "Segmentation fault". Did you find a solution?

[LimHyungTae]

It's hard to reproduce your problem. But, one thing I'm curious is how many correspondences are given to TEASER++?

[arminkmbr]

Hi, thanks for your reply. I'm giving about 600 putative correspondences to Teaser. But the segmentation fault only happens for cases with a high noise bound, e.g. above 1. For low noise bounds, the algorithm works as intended. Funnily enough, I get a segmentation fault and the end of the algorithm even after successful registration.

I'd happily give you more detailed information but im not too sure what info to give.

If i run an example in c++ with extreme high noise bound it takes long but it still works.

I using Win 11 with WSL2 on Ubuntu 20.04 with Python 3.6 and PyCharm. Open3d==0.15.2

[LimHyungTae]

Hmmm it's weird, so you mean you only have a problem when you run TEASER++ on Python? Please use gdb and check the lines where the TEASER++ code makes a breakdown.

[arminkmbr]

Hi @LimHyungTae
yes only if i run the python examples with high noise bound. if i run c++ even with high noise bound it runs perfectly. as im having issues with vs code and such i started using the python implementation.

im running with os.environ["OMPNUMTHREADS"] = "12"
this is the output from gdb:

Starting program: /home/armin/anaconda3/envs/teasertest/bin/python teaserpythonply.py
[Thread debugging using libthreaddb enabled]
Using host libthreaddb library "/lib/x8664-linux-gnu/libthread_db.so.1".
[Detaching after fork from child process 76685]
[Detaching after fork from child process 76686]
[Detaching after fork from child process 76689]
[Detaching after fork from child process 76690]
[Detaching after fork from child process 76691]
[Detaching after fork from child process 76694]
[New Thread 0x7f3509278700 (LWP 76695)]
[New Thread 0x7f34dacfd700 (LWP 76696)]
[New Thread 0x7f34d9714700 (LWP 76697)]
[New Thread 0x7f34d8f13700 (LWP 76698)]
[Thread 0x7f34d8f13700 (LWP 76698) exited]
[New Thread 0x7f34d39f8780 (LWP 76699)]
[New Thread 0x7f34d31f6800 (LWP 76700)]
[New Thread 0x7f34d29f4880 (LWP 76701)]
[New Thread 0x7f34d21f2900 (LWP 76702)]
[New Thread 0x7f34d19f0980 (LWP 76703)]
[New Thread 0x7f34d11eea00 (LWP 76704)]
[New Thread 0x7f34d09eca80 (LWP 76705)]
[New Thread 0x7f34c5ffeb00 (LWP 76706)]
[New Thread 0x7f34c57fcb80 (LWP 76707)]
[New Thread 0x7f34c4ffac00 (LWP 76708)]
[New Thread 0x7f34c47f8c80 (LWP 76709)]
[Detaching after fork from child process 76710]
[Detaching after fork from child process 76712]
[New Thread 0x7f34d8f13700 (LWP 76713)]
[New Thread 0x7f34c1586700 (LWP 76714)]
[New Thread 0x7f34c0d85700 (LWP 76715)]
[New Thread 0x7f34802b2700 (LWP 76716)]
[New Thread 0x7f347dab1700 (LWP 76717)]
[New Thread 0x7f347b2b0700 (LWP 76718)]
[New Thread 0x7f3478aaf700 (LWP 76719)]
[New Thread 0x7f34782ae700 (LWP 76720)]
[New Thread 0x7f3473aad700 (LWP 76721)]
[New Thread 0x7f34712ac700 (LWP 76722)]
[New Thread 0x7f346eaab700 (LWP 76723)]
TEASER++ Python registration example
[Detaching after vfork from child process 76725]
[New Thread 0x7f345a74e700 (LWP 76727)]
[New Thread 0x7f344beb8700 (LWP 76728)]
[New Thread 0x7f344b6b7700 (LWP 76729)]
[New Thread 0x7f344aeb6700 (LWP 76730)]
[New Thread 0x7f344a6b5700 (LWP 76731)]
Starting scale solver (only selecting inliers if scale estimation has been disabled).
Scale estimation complete.
--Type for more, q to quit, c to continue without paging--
Thread 1 "python" received signal SIGSEGV, Segmentation fault.
0x00007f345cc13f10 in teaser::RobustRegistrationSolver::solve(Eigen::Matrix const&, Eigen::Matrix const&) () from /home/armin/PYTEASER/build/teaser/libteaserregistration.so

[LimHyungTae]

How much large noise bound do you use now? Originally, a 'large' noise bound is not intended; it should be no larger than voxel sampling size (if you conduct voxel sampling as a preprocessing)

[arminkmbr]

Well, im trying to register 2 point clouds from different sources. One pcd is highly accurate extracted from a 3D model and the other is noisy/incomplete reconstructed from SLAM. I have extracted the columns from both point clouds and extracted the center point from each column. Now im trying to register these column center points with all-to-all correspondence. It works very well when I use the same cloud for source and target. I did not get it to work for the cross-source case with noise bound above 2.5. I have around 25 column center points on each pcd.

My idea was to test to increase the noise bound to compensate for the inaccuracies from the SLAM column centers...
If I could just get the translation, that would be beneficial and I could do more testing. Maybe i am not not fully understanding the paper yet.

[jingnanshi]

@arminkmbr All-to-all correspondences are going to be difficult for your case with a large number of points. The segfault might be caused by the program using too much memory.

[Guoc0529]

[New Thread 0x7fff53ff7700 (LWP 1431930)]
[New Thread 0x7fff537f6700 (LWP 1431931)]
[New Thread 0x7fff547f8700 (LWP 1431932)]
[New Thread 0x7fff51ff3700 (LWP 1431933)]
[New Thread 0x7fff577fe700 (LWP 1431934)]
[New Thread 0x7fff56ffd700 (LWP 1431935)]
[New Thread 0x7fff567fc700 (LWP 1431936)]
[New Thread 0x7fff55ffb700 (LWP 1431937)]
[New Thread 0x7fff557fa700 (LWP 1431938)]
[New Thread 0x7fff54ff9700 (LWP 1431939)]
[New Thread 0x7fff52ff5700 (LWP 1431940)]
[New Thread 0x7fff527f4700 (LWP 1431941)]
Starting scale solver (only selecting inliers if scale estimation has been disabled).
Scale estimation complete.
Max core number: 2
Num vertices: 53
Max Clique of scale estimation inliers:
Using chain graph for GNC rotation.
Starting rotation solver.
GNC rotation estimation noise bound:0.6
GNC rotation estimation noise bound squared:0.36
GNC-TLS solver terminated due to cost convergence.
Cost diff: 0
Iterations: 5
Rotation estimation complete.
Starting translation solver.
Translation estimation complete.
double free or corruption (out)

Hello, this always appears before each segmentation fault. Do you have any good solutions?

[LimHyungTae]

Did you set OMP_NUM_THREADS=16 in front of your command?