didn’t get a good reconstrcution result by directly using sdf.implicit_surface_to_mesh_directory

[QtEngineer]

Hi, for testing ,I tried to reconstruct objs from abc_train datasets .I only used fuction sdf.implicit_surface_to_mesh_directory with argument query_dist_dir='abc_train\05_query_dist', query_pts_dir='abc_train\05_query_pts', query_grid_resolution=128 sigma=5 certainty_threshold=13. I opened gernerated ply&off file with meshlab , found all of them have many holes ,which quite different from GTmesh . Are there any problems when I run the code?

[ErlerPhilipp]

Which model did you use? Which dataset? Even the full (vanilla) model has some problems with certain objects, i.e. open objects and landscapes. Also, the heuristic and sub-sampling is sensitive to varying point cloud density. Here is an unfortunate example from the Iso-Points paper:

[QtEngineer]

I didnot use any model，only use 05_query_dist and 05_query_pts from ABC_train datasets . Below is the result of 00010006_7e4956ae07e24f6584127385_trimesh_000.ply (generated by function sdf.implicit_surface_to_mesh_directory) I am not sure whether it can generated correct mesh . I found ABC_train dataset doesnot contain 03_mesh and I want to dirrectly get it but seems failed.

[ErlerPhilipp]

Ahh I see. 05_query_dist and 05_query_pts contain only sparsely / randomly sampled training data. It's not a proper representation of the entire shapes and won't produce a good reconstruction.

The GT shapes (03_meshes) should be included in the test sets for download. If you want the original input meshes, you would need to search the ABC dataset. To make things easier, I just uploaded the 03_meshes here: https://www.cg.tuwien.ac.at/research/publications/2020/erler-2020-p2s/erler-2020-p2s-abc_meshes.zip

[QtEngineer]

Thanks a lot!