Triangle in 3D
Closed this issue · 4 comments
Hi, @zexinyang
Excellent work on cloud registration!
I noticed you use Cloud3D in the code instead of Cloud2D. So I wonder would it work well in actual triangles in 3D, for example the vertexes are not strictly corresponded but in most similarity.
Thank you for your codes.
Hi, thanks for your interest in our work!
Our matching algorithm works in 3D, meaning the triangles are constructed in the 3D space (instead of the XOY plane). In other words, stem positions (i.e., the vertices of triangles) differ in height due to the bumpy DTM surface, which forms 3D triangles.
As real-world data contains measurement errors, there are no perfectly corresponding vertices. That’s why we use a distance threshold to determine if two vertices (triangles/graphs constructed from triangle pairs) are matched.
I hope this helps! Let me know if you have any other questions.
Thank you very much, @zexinyang
Instead of stem positions, I actually want to register two point clouds based on the triangles/graph constructed by some keypoints. So far, the results are not satisfied with around 4~7 finally matched vertexes. Would you please give me some advice/insights on this task?
That should be doable. We have also successfully applied our matching method to align cross-source and multi-temporal point clouds in urban areas.
To ensure accurate registration results, I recommended your keypoints should be relatively accurate. The errors of your keypoints (i.e., the distance threshold) should be much smaller than the distances among them. For example, when we were registering forest TLS scans, stem positions have a (probably) centimeter-level accuracy while the distances among them are much larger (around meters). By setting the distance threshold to 5 cm, our method worked perfectly. Similarly, when registering ALS point clouds with photogrammetric ones in urban scenes, our keypoints have an accuracy of decimeters while the average distances among them are more than ten meters. We obtained good registration results by setting the distance threshold to 10-20 cm.
You can also visualize the constructed triangles to see if the K value is large enough - are there corresponding triangles in the overlapping areas? Increasing the K value may help if you would like to align a pair of point clouds with limited overlaps.
For now, you can adjust the distance threshold (edge_diff_) and K value (knn_) for your data from here. I will add setters and comments in the near future =)
Thank you very much for your generous analysis. I'm clear.