Towards benchmarks and a database for context recognition