Scene Understanding Database With 899 Categories

The Scene UNderstanding (SUN) database contains 108,754 images across 899 scene categories, with at least 100 images per category. The images are provided for research purposes in formats including JPG, PNG, or GIF.

Use Cases

• Train a multi-label scene classifier using the 899 distinct category labels.
• Analyze visual feature distributions across the 899 scene categories to study intra-class variation.
• Benchmark scene recognition models on a dataset with a minimum of 100 images per category.
• Develop few-shot learning methods leveraging the structured hierarchy of 899 scene types.

Strengths

• Contains 108,754 total images, providing substantial data for model training.
• Covers 899 distinct scene categories, offering broad coverage of visual environments.
• Ensures a minimum of 100 images per category, supporting reliable per-class analysis.

Limitations

• The specific number of images per category varies, which may lead to class imbalance.
• Image formats include GIF, which may require specific preprocessing for computer vision tasks.
• No column or metadata structure is provided, limiting structured analysis beyond image classification.

Provenance

Source

HuggingFaceM4 via Hugging Face platform.

Freshness

Last updated in July 2022.