csl-aaa/SketchGraph

GitHub repository: https://github.com/csl-aaa/SketchGraphNet

The repository provides scripts and instructions for loading the SketchGraph dataset during training.

Dataset Summary

SketchGraph is a large-scale dataset of graph-structured sketches designed for studying graph-based sketch understanding methods. The dataset contains 3.44 million sketches across 344 semantic categories, with 10,000 samples per category.

Each sketch is converted into a spatiotemporal graph representation with a fixed node count (n=100), obtained via uniform point sampling along drawing strokes. This design allows consistent batch processing during large-scale training.

Two dataset variants are provided:

• Version A: unfiltered sketches directly derived from QuickDraw
• Version R: sketches verified as recognizable by the QuickDraw recognition system

Both variants share identical category sets and preprocessing procedures.

Source Data

SketchGraph is derived from the QuickDraw dataset. The category “line” is excluded because of its trivial topology and minimal stroke complexity.

The remaining 344 categories are processed using balanced sampling to reduce class imbalance.

Dataset Variants

SketchGraph provides two dataset variants:

Version A

• contains all sketches derived from QuickDraw
• includes sketches with varying drawing quality and structural consistency

Version R

• contains sketches verified by the QuickDraw recognition system
• exhibits more consistent structural patterns

Both variants include 10,000 samples per category.

Dataset Size

Property	Value
Categories	344
Samples per category	10,000
Total samples	3,440,000
Nodes per graph	100

Data Storage Format

SketchGraph is stored using binary serialization with offset-based indexing.

This storage format enables:

• parallel random access
• efficient batch loading
• high-throughput I/O during training