datasets

Why we included this project

Most vision research leans on cleanly labeled academic sets, which say little about how images actually get used in the wild. The Unsplash dataset goes the other way: 7.4 million real photos paired with 1.6 million keywords and over 160 million search records, so the labels carry genuine human intent rather than tidy ground truth. That makes it a workable base for training classifiers, semantic search, and image-text retrieval on messy real-world data. A Lite version, roughly 25,000 photos with the same fields, is fine for testing an approach and carries a commercial-friendly license, while the full dataset requires a non-commercial request.