Quick, Draw! Subset: 414,000 Images for Neural Network Pre-training

414,000 drawings from the Quick, Draw! dataset, processed into 256x256 pixel images. Juan Guerrero Martín curated this subset, selecting 1,200 images for each of 345 classes and splitting them into training, validation, and test sets. The dataset was last updated on October 14, 2025.

Use Cases

• Pre-training neural networks for image recognition based on the large collection of 345 drawing classes.
• Benchmarking model performance on sketch data using the provided training, validation, and test splits.
• Developing classifiers for hand-drawn objects based on the standardized 256x256 pixel format.

Strengths

• Contains 414,000 images, providing substantial volume for model training.
• Includes 345 distinct drawing classes, offering broad categorical coverage.
• Data is pre-split into 289,800 training, 62,100 validation, and 62,100 test images.
• All images are standardized to a 256x256 pixel resolution.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset's connection to the Rey-Osterrieth Complex Figure test mentioned in the project description is not detailed.

Provenance

Source

Subset derived from the original Quick, Draw! dataset.

Collection Method

Images were downloaded in binary format, selected, converted from vector to pixel format, and resized.

Time Range

null

Freshness

Last updated 2025-10 14 21:38:28; freshness should be verified.

Geography

null