Describable Textures Dataset with 47 Human-Perception Categories

The DTD dataset contains 5,640 texture images organized into 47 categories based on human perception, with 120 images per category. Images were collected from Google and Flickr and annotated via Amazon Mechanical Turk, with each image assigned a main category and joint attributes. Image sizes range from 300x300 to 640x640 pixels.

Use Cases

• Train a multi-label classifier to predict the main category and joint attributes for texture images.
• Fine-tune a vision model on the 47 texture categories for material recognition tasks.
• Analyze the co-occurrence patterns of joint attributes across the 5,640 images.
• Use the 120 images per category for texture synthesis or data augmentation studies.

Strengths

• Contains 5,640 images, providing a substantial collection for texture analysis.
• Organized into 47 distinct, human-perception-inspired categories for structured learning.
• Each category has a balanced 120 images, reducing class imbalance concerns.
• Images contain at least 90% surface area representing the assigned category attribute, ensuring label relevance.

Limitations

• Image sizes vary between 300x300 and 640x640 pixels, requiring resizing for uniform model input.
• Annotations were crowdsourced via Amazon Mechanical Turk, which may introduce label noise.
• The dataset's collection from Google and Flickr may reflect search engine and platform biases in image representation.

Provenance

Source

Images collected from Google and Flickr.

Collection Method

Images gathered by entering proposed attributes as search queries and annotated via Amazon Mechanical Turk.

Freshness

Last updated on 2022-06-22.