chibifire/datasource-garmentcodedata-v2-parquet

GarmentCodeData v2: patterns, meshes and labels

Sewing patterns, draped garment meshes and panel labels from GarmentCodeData v2, repacked as ZSTD Parquet. Each garment design is fitted to and draped on a neutral body and on a sampled body.

What it is for

A sewing pattern paired with the shape it takes when sewn and draped. Use it to start a garment from a real pattern, and to compare a cloth solver's drape against the upstream simulator on the same pattern.

Provenance

Upstream: GarmentCodeData v2, Korosteleva et al., ETH Zurich, 2024 (doi 10.3929/ethz-b-000690432), CC BY 4.0, read from the ETH libdrive share named in that record. Each batch's data.tar.gz was streamed and its files were copied byte for byte. The sha256 column is the hash of each file as read from the tar.

Kept per garment (kind column, upstream file suffix):

• design_params (_design_params.yaml): the design parameters that made the pattern.
• body_measurements (_body_measurements.yaml): the measurements the pattern was cut for.
• specification (_specification.json): the sewing pattern.
• pattern_svg (_pattern.svg): the pattern panels as vector art.
• box_mesh (_boxmesh.ply): the panels placed around the body before draping.
• sim_mesh (_sim.ply): the draped garment.
• segmentation (_sim_segmentation.txt): the panel label of each vertex.
• vertex_labels (_vertex_labels.yaml): waist, collar and strapless-top vertex labels.

Not kept: the PNG renders, the PNG pattern, the texture, the rest edge-length pickle, and the 5,000 sampled body shapes. upstream/ holds the official split, the documentation, the neutral body and each batch's dataset_properties file, unchanged.

Fabric per batch, from the upstream documentation: batches 0-6 and 21-25 use bending 1.0, 7-13 and 26-30 use 100, 14-20 and 31-35 use 500.

Use

One row per file: batch, body (default or random), element, kind, data, sha256. A garment is the set of rows sharing batch, body and element.

import pyarrow.parquet as pq
t = pq.read_table("default_body/batch-00.parquet", filters=[("kind", "=", "specification")])

The official split lists garments_5000_<batch>/<body>_body/<element> paths in upstream/GarmentCodeData_v2_official_train_valid_test_data_split.json.

PLY meshes keep their UVs; per the upstream documentation, load them with trimesh.load_mesh(path, process=False). trimesh inserts a duplicate vertex at each UV seam right after the original, so map ids back before reading the segmentation or vertex labels.

Licence

CC BY 4.0, as the upstream record states. Cite the upstream paper and dataset (see CITATION.cff). The upstream body model was fitted to the European subset of CAESAR; its authors state it covers healthy European and North American adults.