Skip to content

DATA.md

Latest commit

 

History

History
61 lines (41 loc) · 2.34 KB

DATA.md

File metadata and controls

61 lines (41 loc) · 2.34 KB

Data Preprocessing

We provide scripts that preprocess raw data (i.e., 3D articulated assets) in urdf or usd format into training/evaluation formats.

URDF

To preprocess a urdf asset, use the process_urdf.py script under particulate/data:

python -m particulate.data.process_urdf path/to/urdf/file.urdf path/to/output/folder

The script will save an OBJ file original.obj under the output folder, together with meta files meta.npz, link_axes_plucker.npz, and link_range.npz which store the part and articulation related metadata.

USD

To preprocess a usd asset, use the process_usd.py script under particulate/data:

python -m particulate.data.process_usd path/to/usd/file.usd path/to/output/folder

The script will save files under the output folder in the same format as the one for urdf assets: an OBJ file original.obj under the output folder, together with meta files meta.npz, link_axes_plucker.npz, and link_range.npz which store the part and articulation related metadata.

Training Data

For training, we cache points and relevant metadata for efficient data loading. To do this for a preprocessed output folder, use the cache_points.py script under particulate/data:

python -m particulate.data.cache_points --root path/to/the/preprocessed/folder --output_path path/to/the/output/npz/file.npz --num_points 40000 --ratio_sharp 0.5

which caches a training datum containing the following attributes to the output path:

points
normals
point_to_bone
point_from_sharp
bone_structure
link_axes_plucker
link_range

The arg --ratio_sharp dictates the percentage of total points which are sampled from sharp edges.

Evaluation Data

To convert Training Data to Evaluation Data, run:

python -m particulate.convert_train_to_eval --input_file /path/to/the/preprocessed/file --output_file path/to/the/output/npz/file.npz

For evaluation, we cache points and relevant metadata for efficient data loading. In our evaluation, we cache 100k points uniformly sampled from the surface of each mesh. To do this for a preprocessed output folder, use the cache_gt.py script under particulate/data:

python -m particulate.data.cache_gt  --root_dir /path/to/directory/of/preprocessed/assets/ --output_dir /path/to/save/cached/ground/truths/