SVHN: 600,000+ Real-World Street View House Number Images

SVHN contains over 600,000 labeled images of house numbers captured from Google Street View, developed by researchers at Stanford University. It serves as a real-world benchmark for digit recognition that incorporates natural variations in lighting, font, and background noise.

Use Cases

• Image classification of digits 0-9 in natural environments
• Object detection for locating house numbers in street-level imagery
• Benchmarking model performance against real-world noise and distortions

Strengths

• Over 600,000 labeled digit images
• Expert-generated annotations for ground truth
• High real-world variability in lighting, color, and geometric distortions

Limitations

• Geographic bias toward United States street imagery
• Restricted to numeric digits 0-9, excluding alphabetical characters

Provenance

Source

Stanford University (UFLDL)

Collection Method

Scraped from Google Street View and annotated by experts and machines.

Freshness

Last updated August 2024 on the Hugging Face platform.

Geography

United States