Problems with registration

[matfcp]

Hi! First of all, congratulations for the fantastic work!

I'm having issues when trying to register two point clouds from my own dataset. It consists of a set of virtual pottery sherds obtained from breaking a virtual pottery vessel in Blender. Here are the pictures from the two sherds' clouds I'm using:

I ran the "assemble_fragments.py" with default parameters and got the following results:

You can notice that the registration is not good since pottery sherds should fit with their borders (breaking faces) side by side.

Here are the border segmentation results ("col_borders_obj1_challenge.ply" and "col_borders_obj2_challenge.ply"):

As you can see, not only the borders are being segmented, but parts of the sherd's body are being classified as borders as well.

Is there a way to improve the results? Unfortunately, this is not a well behaved dataset, so the shapes of sherds and borders are random.

Thanks a lot in advance!

[freerafiki]

Hi - sorry for the delay, I did not manage to stay tuned for some time. It seems from looking at the result that the "red" part (the borders) are too "thick", if this makes sense. Because the density of the cloud is most likely different from the previously used data, my feeling is that the dilation part is creating a huge part of borders, so I think you could set a different parameters to obtain smaller "red" part and larger "green" part (which should be the part which fits, if I see it correctly?).

The best would be to play a little bit with the parameters (check here for better explanation) and see if you can get a better results for the segmentation. If segmentation is correct, the matching part may work as well.

Do you have any further result btw?

[matfcp]

Hi! No problem, thank you so much for your thoughtful response!

I don't have any more results yet, but I'll try playing with the parameters you just pointed, so I can come back to tell what I get.

[freerafiki]

Thanks! I see that few people have pointed out that the accuracy oscillates more than expected, but I also see interests on the topic (because I think not so many "easy to use" methods are available) so I actually hope we will be able to improve the robustness of the breaking curve detection! Thanks if you will help with your results!

[matfcp]

Indeed, your method is very straightforward, making it really appealing. In fact, this approach really interests me since I'm working with pottery vessel reconstruction with the aid of deep neural networks. My work focused on estimating the relative position of pottery sherds, which can (maybe) give us a good first guess about the initial position of the point cloud of a sherd to then perform the registration. I'm currently searching for registration methods that can help me with this next stage of the research, which consists of the registration itself.

[matfcp]

Hi again! I tried playing with the parameters as you suggested, but things didn't get much better. I varied the parameters like

to = 200
tb = 0.05
dil = 0.008
thre = 0.88

or

to = 100
tb = 0.1
dil = 0.008
thre = 0.88

but depending on the values I use, the segmentation results improve for one fragment but gets poorer for the other. I'd appreciate some insights on that, if possible. Thanks for your attention!