SIFT1B Dataset

Overview

SIFT1B (also known as BigANN or ANN_SIFT1B) represents the 128-dimensional SIFT (Scale-Invariant Feature Transform) descriptors of one billion images. Released in September 2010, it remains a fundamental benchmark for large-scale vector search evaluation.

Dataset Characteristics

• Size: 1 billion vectors
• Dimensions: 128-dimensional SIFT descriptors
• Source: Image feature descriptors
• Format: High-dimensional vectors suitable for similarity search

Significance for Evaluation

SIFT1B plays a critical role in evaluating vector search algorithms by providing:

• Consistent, reproducible foundation for comparison
• Billion-scale testing to stress-test ANN algorithms
• Pre-processed data with known characteristics
• Industry-standard benchmark for performance claims

Related Datasets

• SIFT1M: 1 million SIFT descriptors (smaller version for initial testing)
• GIST1M: 960-dimensional GIST descriptors
• Deep1B: 1 billion deep learning features

Dataset Access

Laurent Amsaleg (CNRS/IRISA) and Hervé Jégou (Facebook AI Research) have waived all copyright and related rights. Datasets can be downloaded from http://corpus-texmex.irisa.fr/

For downloading BIGANN, using Axel is recommended for faster downloads.

Usage in Research

SIFT1B is extensively used in:

• ANN algorithm benchmarking
• Vector database performance evaluation
• Scalability testing
• Algorithm comparison studies