What it is
Benchmark for extremely hard image matching with 1,000 hand-annotated pairs and a simple Python evaluation API. Evaluate local feature matchers on 1,000 extremely difficult hand-annotated image pairs through a simple benchmark API, with built-in data download and baseline demos.
Key results
- Covers 1,000 hand-annotated image pairs
- Splits into 900 test pairs and 100 validation pairs
- Benchmark auto-downloads the 660MB dataset
Shows that current local feature matchers need evaluation on much harder real-world pairings than standard benchmarks.