HMDB: A large video database for human motion recognition

Motivated by the gap between rich static image datasets and the small, tightly controlled human action datasets available at the time, the authors assembled HMDB, the largest action video database collected to date. It contains 51 distinct action categories and around 7,000 manually annotated clips gathered from varied sources, including digitized movies and YouTube videos.

Using this benchmark, they evaluated two representative computer vision systems for action recognition and examined how their performance held up under challenging conditions such as camera motion, changing viewpoint, low video quality, and occlusion. By offering far more categories and realistic footage than prior benchmarks whose performance was near ceiling, the database provided a harder and more realistic testbed for advancing video action recognition.