search

ErlerPhilipp / points2surf

Learning Implicit Surfaces from Point Clouds (ECCV 2020)
https://www.cg.tuwien.ac.at/research/publications/2020/erler-2020-p2s/
MIT License
443 stars 46 forks source link

Cannot generate all Famous dataset for training #23

Closed QtEngineer closed 1 year ago

QtEngineer commented 2 years ago

I tried to use make_dataset.py to make famous_origin dataset for training but only got 3 files : WARNING:only some output files are missing: ['datasets/famous_original/04_pts_vis/Utah_teapot_solid.xyz'] ..., flower.xyz ..., Cup34.xyz

QtEngineer commented 2 years ago

both error accured when I tried to preproces ShapeNet with about 1000abjs,but only got 50

ErlerPhilipp commented 2 years ago

make_dataset.py has a parameter "num_max_faces", which filters all input meshes with more faces. you can set this to "None" to allow all meshes. the reason for this is that the SDF calculation in Trimesh takes very long, depending on the mesh complexity. there are other requirements like they must be watertight and 2-manifold.

most ShapeNet objects have a very bad quality. when i tried, i couldn't even calculate vertex normals because of all the non-manifold pieces. you need either to clean the shapenet objects or use a different SDF calculation.

QtEngineer commented 2 years ago

thanks a lot！