HMDB-51 Dataset

With almost 1B online videos viewed across the world everyday, an emerging new frontier in computer vision research is that of recognition and search abilities in videos. While a great deal of effort has been directed towards the collection and annotation of large scalable datasets of static image, the efforts towards human action datasets relatively lack behind. HMDB is collected from a variety of sources, mostly movies, but also a small proportion has been obtained from open source databases such as the Prelinger archive, YouTube, and Google videos. This dataset contains ~7k clips divided into 51 action categories, each containing over 101 clips.


Download Instructions

1 )Go to:
2) Download the Video Database Files, STIP Features and Three splits files