Textures-DTD: 1,880 Preprocessed Texture Images for Few-Shot Learning

1,880 square images of textures, preprocessed from the Describable Textures Dataset for the Meta-Album meta-dataset. The data consists of 47 texture categories, with images cropped and resized to 128x128 pixels. Ihsan Ullah created this version in March 2022, based on original data collected by researchers at the University of Oxford from Google and Flickr.

Use Cases

• Benchmarking few-shot image classification models based on the 47 texture categories.
• Training texture recognition models for manufacturing quality inspection based on the preprocessed 128x128 images.
• Developing data augmentation strategies for small-scale image datasets based on the curated texture samples.

Strengths

• Contains 1,880 images across 47 distinct texture classes, providing a multi-class benchmark.
• Images are consistently preprocessed to a uniform 128x128 pixel size, which may simplify model input pipelines.
• Clear provenance from the well-known Describable Textures Dataset (DTD) and associated academic paper.

Limitations

• Description metadata is limited; actual data quality, such as label accuracy or image clarity, requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is a mini subset of the original DTD, which contained 5,640 images, indicating a reduced scale.

Provenance

Source

Describable Textures Dataset (DTD), University of Oxford, England.

Collection Method

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