NDT_CUDA get nan result.

[lnexenl]

My system info:

pcl 1.12.1 eigen 3.4.0 CUDA 11.6 cudnn 8.3.2 gcc 11.2.0

My cmake argument:

cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_STANDARD=17 -DCMAKE_INSTALL_PREFIX=/opt/ros/noetic -DBUILD_VGICP_CUDA=ON

---

Hi I am currently using fast_gicp to do some prior locating work. However, I found that when using NDT_CUDA to register my pointclouds, I get nan results:

target:67804[pts] source:14093[pts]
--- pcl_gicp ---
single:485.159[msec]     
0.999971  0.000591522    0.0076433    -0.085595
 -0.00058857            1 -0.000388423    0.0937259
 -0.00764352  0.000383913     0.999971      2.30524
           0            0            0            1
--- pcl_ndt ---
single:61.4179[msec]     
0.999999   0.00120411  4.85903e-05   -0.0260552
 -0.00120416     0.999999   0.00107856    0.0946873
-4.72916e-05  -0.00107862     0.999999   -0.0187839
           0            0            0            1
--- fgicp_st ---
single:348.614[msec]     
0.999985  -0.00105683   0.00540941   -0.0871596
  0.00105919     0.999999 -0.000433943   -0.0925524
 -0.00540895  0.000439666     0.999985      2.18176
           0            0            0            1
--- fgicp_mt ---
single:372.079[msec]     
0.999985  -0.00105867   0.00540852    -0.087245
  0.00106108     0.999999 -0.000443048   -0.0925688
 -0.00540805   0.00044878     0.999985       2.1817
           0            0            0            1
--- vgicp_st ---
single:448.121[msec]      
0.99999   -0.0014508   0.00434923   -0.0833737
  0.00145268     0.999999 -0.000428907   -0.0780743
  -0.0043486  0.000435221      0.99999      2.12142
           0            0            0            1
--- vgicp_mt ---
single:549.662[msec]      
0.99999   -0.0014508   0.00434923   -0.0833737
  0.00145268     0.999999 -0.000428907   -0.0780743
  -0.0043486  0.000435221      0.99999      2.12142
           0            0            0            1
--- ndt_cuda (P2D) ---
single:49.4802[msec] 
nan nan nan nan
nan nan nan nan
nan nan nan nan
  0   0   0   1
--- ndt_cuda (D2D) ---
single:51.1368[msec] 
nan nan nan nan
nan nan nan nan
nan nan nan nan
  0   0   0   1
--- vgicp_cuda (parallel_kdtree) ---
single:96.7676[msec]     
 0.99999  -0.00145048   0.00434936   -0.0834129
  0.00145235     0.999999 -0.000428305    -0.078238
 -0.00434873  0.000434617      0.99999      2.12144
           0            0            0            1
--- vgicp_cuda (gpu_bruteforce) ---
single:157.454[msec]     
0.999938  -0.00771049   0.00800458    0.0143744
  0.00770667      0.99997   0.00050875      -3.1288
 -0.00800827 -0.000447029     0.999968      2.43394
           0            0            0            1
--- vgicp_cuda (gpu_rbf_kernel) ---
single:62.088[msec]     
0.999989  -0.00153041   0.00450023    -0.119624
  0.00153244     0.999999 -0.000448908   -0.0212703
 -0.00449954  0.000455799      0.99999      2.12726
           0            0            0            1

Registrations give me some non-nan results except NDT_CUDA. Using the pointclouds provided by the repo, the NDT_CUDA algorithm still produces non-nan result.

The two pointclouds used by me: pcd.zip

[koide3]

It seems the source is too apart from the target, and optimization is broken because there is completely no overlap. I'm not sure why other methods don't get corrupted, but I think any method can get broken in this setting.

[lnexenl]

It seems the source is too apart from the target, and optimization is broken because there is completely no overlap. I'm not sure why other methods don't get corrupted, but I think any method can get broken in this setting.

However on I found that align pointclouds provided by this repo also produces same result. I don't know whether this relates to PCL version, cause I found that current code in the repo is a bit incompatible with pcl v1.12.

[whuzs]

Have you solved this problem? I suspect it is related to the versions of eigen and cuda

[lnexenl]

Have you solved this problem? I suspect it is related to the versions of eigen and cuda

Yes, I found that PCL 1.12 is not compatible with this library. I downgrade PCL to 1.10 and everything works well. I think there might be some bugs with PCL 1.12. Besides, I am using archlinux.

[whuzs]

thanks, i am using orin(aarch64), cuda11.4,pcl1.10.0, it still get nan result(ndt cuda)...

[dch4o]

thanks, i am using orin(aarch64), cuda11.4,pcl1.10.0, it still get nan result(ndt cuda)...

Hi @whuzs. Did you solved the problem? I am currently using PCL 1.11.1 and Jetpack r35.1, but no problem using this library.