3DMatch Dataset - Wrong Mathces

[yuvalH9]

Hi, I was wondering if you validated the correspondences between src and tgt over the trainset. If I understand correctly, the point clouds given in the .pkl files are already aligned and the correspondences are for the aligned point clouds given as an ordered list.

I tried to visualize myself some random corresponding points and it seems that the points are not matches at all.

I attached an example for the first matching pair (visualized by Matlab) with 10 random matches

[XuyangBai]

Hi, I didn't see the example. Could you specify which point clouds pair you are using so that I can see it on my environment

[yuvalH9]

Hi @XuyangBai , I had a visualization problem and after fixing that the correspondence seems to be correct.

By the way, if I want to create my own GT correspondence for the train and validation dataset, how can I do that? Maybe you can upload the GT transformations for the train and validation datasets (similar to the format of the test dataset) so we can create our own GT correspondence?

Thanks! Yuval

[XuyangBai]

Hi, You can download the raw dataset and run https://github.com/XuyangBai/PPF-FoldNet/blob/master/script/fuse_fragments_3DMatch.py to compute the point clouds and the corresponding poses, from which you can compute the GT transformations.