Textures-DTD: 800 Images of 20 Textures for Few-Shot Learning

800 square images of 20 different texture classes, resized to 128x128 pixels. The data originates from the Describable Textures Dataset (DTD) collected by researchers at the University of Oxford and was preprocessed for the Meta-Album meta-dataset in March 2022. Images were sourced from Google and Flickr and annotated via Amazon Mechanical Turk.

Use Cases

• Few-shot image classification based on 20 texture classes.
• Texture recognition model training using 128x128 pixel images.
• Benchmarking computer vision algorithms on a standardized, preprocessed subset of a known texture dataset.

Strengths

• Contains 800 images across 20 distinct texture classes.
• Images are uniformly preprocessed to a square 128x128 pixel format.
• Has a clear provenance from the academic Describable Textures Dataset (DTD).

Limitations

• Row count is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is a micro subset, representing a fraction of the original DTD's 5,640 images.

Provenance

Source

Describable Textures Dataset (DTD), University of Oxford.

Collection Method

Images collected from Google and Flickr, annotated via Amazon Mechanical Turk, and preprocessed (cropped and resized) for Meta-Album.