README.md

metadata

pretty_name: Shared-Class Sketch
task_categories:
  - image-classification
tags:
  - synthetic
  - sketch
  - compositionality
  - hierarchical-representation
size_categories:
  - 10K<n<100K

Shared-Class Sketch

Shared-Class Sketch is a deterministic synthetic corpus for testing whether a model can represent sketch categories through reusable primitives and parts. Every image is a 128 x 128 binary raster with white strokes on a black background. No real sketch raster or stroke is copied into the dataset.

Classes and structural variation

The ten semantic classes are bicycle, truck, binoculars, face, windmill, birthday_cake, butterfly, chandelier, flower, and guitar. Each class contains five balanced structural families:

• bicycle: diamond, open frame, beam, step-through, and compact;
• truck: box truck, pickup, dump truck, tanker, and flatbed;
• binoculars: porro prism, roof prism, compact, field glass, and opera glasses;
• face: round smile, oval neutral, square jaw, long surprised, and wide grin;
• windmill: four sail, six sail, Dutch, farm wheel, and pinwheel;
• birthday cake: tiered round, layer cake, single candle, many candles, and cupcake;
• butterfly: rounded four-wing, angular swallowtail, monarch spotted, side-swept, and compact moth;
• chandelier: three-arm, five-arm, pendant, tiered crystal, and candle wheel;
• flower: five-petal, many-petal daisy, tulip, bell flower, and sunflower;
• guitar: classical, dreadnought, electric cutaway, triangular, and double cutaway.

Structural-family names, primitive counts, geometry profiles, and drawing-style profiles are audit metadata. The intended model supervision consists only of the semantic class and object center; these audit fields are not intended as training targets.

Splits

Split	Images	Purpose
train	20,480	Single-object model training
val	2,560	Single-object checkpoint selection
heldout	2,560	Unseen elongated plus loose-overdrawn factor combination
composition	4,320	Forty-five unseen unordered two-class pairs, 96 images per pair

Archetype identifiers are disjoint across all splits. Training contains the elongated geometry factor and the loose-overdrawn style factor separately, but never their joint combination. The composition split contains no paired scene used for optimization.

Files

• train/, val/, and heldout/ contain class-named image folders.
• composition/ contains one folder for each unordered class pair.
• metadata/*.jsonl contains class and object-center records used by the HDD experiment.
• records.jsonl contains complete audit provenance for every image.
• manifest.json contains counts, checksums, source-disjointness, topology coverage, foreground/border audits, and binary-image uniqueness statistics.
• generation_config.json is the strict deterministic generation recipe.
• The PDF previews separately visualize training topology diversity, the held-out factor combination, and unseen two-object compositions.

Image-folder loaders can infer the single-object class labels directly from the directory names. The experiment-specific JSONL metadata is only needed when object-center supervision is desired.

Integrity and provenance

All 29,920 stored rasters are binary and every split is duplicate-free under a SHA-256 hash of the raster bytes. The generation configuration SHA-256 is 9a46c478594748c3a7c84ce971bf3ef4d79914cebfa4581a34fba6c5eab1f24c; the complete records file SHA-256 is 2626c48df7a5d1fd31388bee106764b516de68216424d9e466eda1cfaf75a8c4.

Aggregate style and abstraction statistics were calibrated against the ten corresponding SlowSketch categories. SlowSketch remains an external held-out reference: no SlowSketch sample path, raster, or stroke initializes a synthetic sample.

Generation

From the Hierarchical Detector-Decoder repository root:

PYTHONPATH=.:src python scripts/create_shared_class_sketch_dataset.py \
  --config_json configs/shared_class_sketch_dataset_10class.json

The generator is deterministic and refuses to overwrite an existing output directory unless overwrite is explicitly enabled.