Sketch Datasets

Overview

SketchKit provides a variety of sketch-based datasets for research on sketch recognition, retrieval, and generation, including most of current popular datasets.

All datasets share a unified interface, so you can load them, search sketches by metadata, or convert sketches into the standard Sketch format easily.

Metadata

Each dataset object provides an attribute items_metadata, pandas.DataFrame, which contains useful information for searching and filtering sketches.

items_metadata is your main tool to search, filter, and batch load sketches efficiently.

  • id: unique identifier of each sketch

  • category: sketch category (e.g., “cat”, “dog”, “chair”)

  • split: dataset split (e.g., “train”, “test”, “val”), available in some datasets

Key concepts of dataset parameters

root: Datatype: str, Default: "~/.cache/sketchkit/datasets"

Directory where the dataset is stored or downloaded.
If the dataset is not present locally, it will be automatically downloaded to this path.
root dictionary can be costumed, if not, it will store datasets in the default path.

Example: dataset = QuickDraw(root=tmpdir)

load_all: Datatype: bool, Default: False

Whether to load all sketches in the dataset into memory at once.

  • False: Load metadata only, sketch can be loaded on demand via metadata.

  • True: Load the entire dataset into memory, having faster access but higher RAM usage.

Example: dataset = QuickDraw(load_all=True)

cislab_source: Datatype: bool, Default: False

Whether to download the dataset from the CIS LAB mirror site or original download link.

  • False: Use default source (e.g., Github/official download link) to download dataset.

  • True: Apply CIS LAB mirror site to download dataset.

Example: dataset = QuickDraw(cislab_source=True)

download: Datatype: bool, Default: True

Automatically download the dataset if it is missing locally.

  • True: As description.

  • False: Not recommended. You need to provide the dataset manually, otherwise an error is raised.

split: Datatype: str, Default: "train"

Select which split of the dataset to load. (e.g., “train”, “test”, “val”)

Example: cats = dataset.items_metadata(dataset.items_metadata[“split”] == “train”)

category: Datatype: list[str], Default: None

Load only specific categories of sketches. (e.g., “cat”, “dog”, “chair”)

Example: cats = dataset.items_metadata(dataset.items_metadata[“category”] == “cat”)

Tip

  • load_all should be chosen based on dataset size and memory availability: use False for large datasets.

  • split and categories can be combined to load subset you need.

  • cislab_source is applicable to most mainstream dataset.

Loading a dataset

# Default load (metadata only)
dataset = QuickDraw()

# Load all data into memory at once
dataset = QuickDraw(load_all=True)

# Download from CISLAB CDN
dataset = QuickDraw(root=tmpdir, cislab_source=True)

Searching sketches by metadata

# Search items using metadata
cats = dataset.items_metadata[
    (dataset.items_metadata["category"] == "cat") &
    (dataset.items_metadata["split"] == "train")
]

# Load sketches based on metadata
cats_sketch = [dataset[row.id] for _, row in cats[:100].iterrows()]

Available Datasets

The following are the built-in sketch datasets currently available in SketchKit:

hzySketch([root, load_all, cislab_source, ...])

The hzy dataset, which contains drawing process for high-quality anime line arts.

QuickDraw([root, load_all, cislab_source, ...])

QuickDraw dataset loader and interface.

ControlSketch([root, load_all, ...])

ControlSketch dataset loader (SketchDataset-style).

TUBerlin([root, load_all, cislab_source, ...])

The TU-Berlin dataset contains vector sketches represented with cubic Bézier curves across 250 categories.

TracingVsFreehand([root, load_all, ...])

Tracing-vs-Freehand dataset loader.

OpenSketch([root, load_all, cislab_source, ...])

The OpenSketch dataset contains vector sketches represented with polyline curves.

SketchXPRIS([root, load_all, cislab_source, ...])

SketchX-PRIS-Dataset loader and interface.

Sketchy([root, load_all, cislab_source, ...])

The Sketchy Database (https://sketchy.eye.gatech.edu/), SVG subset.

PhotoSketching([root, load_all, ...])

The PhotoSketching dataset loader.

GMUSketchCleanup([root, load_all, ...])

GMU Rough Sketch Cleanup dataset (SVG parsing version).

Utility Scripts

sketchkit.datasets.fscoco

sketchkit.datasets.differsketching