We introduced this dataset in Points2Surf, a method that turns point clouds into meshes.

It consists of objects from the ABC Dataset, a collection of Famous meshes and objects from Thingi10k. These are mostly single objects per file, sometimes a couple of disconnected objects. Objects from the ABC Dataset are CAD-models, the others are mostly statues with organic structures.

We created realistic point clouds using a simulated time-of-flight sensor from BlenSor. The point clouds have typical artifacts like noise and scan shadows.

Finally, we created training data consisting of randomly sampled query points with their ground-truth signed distance. The query points are 50% uniformly distributed in the unit cube and 50% near the surface with some random offset.

The training set consists of 4950 ABC objects with varying number of scans and noise strength. The validation sets are the same as the test set. The ABC test sets contain 100 objects, Famous 22 and Thingi10k 100. The test set variants are as follows: (1) ABC var (like training set), no noise, strong noise; (2) Famous no noise, medium noise, strong noise, sparse, dense scans; (3) Thingi10k no noise, medium noise, strong noise, sparse, dense scans