PartRetrievalAndAssembly
Required libraries:
joblib matplotlib pillow networkx scipy scikit-learn torch pytorch3d plotly pandas pyrender
You may also need to build mesh_contain by:
(1) cd src/mesh_contain
(2) python setup.py build_ext --inplace
this point inside mesh check code is from https://github.com/autonomousvision/occupancy_networks/tree/master/im2mesh/utils/libmesh
Prepare the data --------------------------- :
Please first download a processed dataset containing shapes and parts from PartNet Faucet category:
https://drive.google.com/file/d/1r4mPxUJfxwv-9SpbMXLO287Vjuoqfxdc/view?usp=sharing
Create a folder PartRetrievalAndAssembly/data, put and unzip the download file in PartRetrievalAndAssembly/data
The data is organized as:
1: "final_shapes", each folder in final_shapes corresponds to one shape, the shape is represented as (1) mesh(trimesh.Trimesh) (2) surface point cloud (3) volumetric point cloud. All shapes are zero centered, upward aligned and rotated to align its symmetrical plane with plane with normal (1, 0, 0)
2: "final_parts", each folder in final_parts corresponds to one shape, then each sub folder within each folder corresponds to one part of this shape. the part is represented as (1) mesh(trimesh.Trimesh) (2) surface point cloud (3) volumetric point cloud. All parts are zero centered, upward aligned and rotated to align its bbox standard world coordinates
IMPORTANT NOTES: If you use your data, please make sure they follow the same assumption : all whole shape geometries are upward aligned to (0, 1, 0), and the symmetry planes of all whole shape geometries have normal (1, 0, 0). One option is to use the function : align_shapes_by_sym_plane(meshes) we provided in pre_align_by_sym.py, the input is a list of trimesh.Trimesh objects.
Run the demo --------------------------------- :
cd src/demo
. run_faucet.sh
The results will be stored in src/demo/ours
The config.py file controls all hyperparameters
Additional notes ----------------------------- :
IMPORTANT NOTES: If you use your data, please make sure they follow the same assumption : all whole shape geometries are upward aligned to (0, 1, 0), and the symmetry planes of all whole shape geometries have normal (1, 0, 0). One option is to use the function : align_shapes_by_sym_plane(meshes) we provided in pre_align_by_sym.py, the input is a list of trimesh.Trimesh objects.