for the past few weeks I've been working on a project regarding 3D object recognition. I'm quite new to point cloud processing and PCL so I've tried the tutorial for object recognition based on correspondence grouping (https://pcl.readthedocs.io/projects/tutorials/en/master/correspondence_grouping.html#correspondence-grouping). I am using voxel grids as keypoints, SHOT descriptors and Hough Voting as correspondence grouping method. I am working with my own point clouds and I've been playing with the parameters for quite some time now. For some of my point clouds the results are pretty good. I've uploaded the used model (Model.pcd) and two scenes (Scene1.pcd, Scene1.png and Scene2.pcd, Scene2.png) and their corresponding visual results (Recognition1(_View2).png and Recognition2(_View2).png) in the following zip-file: https://www.mediafire.com/file/ice0d9qrjml6xik/3D-Object-Recognition.zip/file
The model and the scene point clouds have no color, for better visualization they're shown in a polygon mesh in the png-files.
Model
Scene 2
This is the result for scene 2 (scene cloud is green and found model is red).
Scene 1
While the result for scene 2 is quite good it gets worse for scene 1 applying the same parameters. I can now change the parameters so that I receive a more precise result for that second scene. But at the same time the result will get worse for the first example. So I didn't find parameters yet that allow good object recognition for all of my scenes. I have to change the parameters for every scene individually. In addition to that the results will differ when I rotate the model or when I change the models position. So this toolchain seems quite unstable to me.
Do you have any ideas on how to improve the object recognition. Is it just that I didn't find the right parameters yet? Could I add something to my toolchain or change something? Are there other algorithms available in PCL that could provide better results? Are my point clouds not well suited for this task?
Cheers :)
