VMASS: Massive Dataset of Multi-camera Video for Learning, Classification and Recognition of Human Actions

Abstract

Expansion of capabilities of intelligent surveillance systems and research in human motion analysis requires massive amounts of video data for training of learning methods and classifiers and for testing the solutions under realistic conditions. While there are many publicly available video sequences which are meant for training and testing, the existing video datasets are not adequate for real world problems, due to low realism of scenes and acted out human behaviors, relatively small sizes of datasets, low resolution and sometimes low quality of video.

This article presents VMASS, a dataset of large volume high definition video sequences, which is continuously updated by data acquisition from multiple cameras monitoring urban areas of high activity. The VMASS dataset is described along with the acquisition and continuous updating processes and compared to other available video datasets of similar purpose. Also described is the sequence annotation process. The amount of video data collected so far exceeds 4000 hours, 540 million frames and 2 million recorded events, with 3500 events annotated manually using about 150 event types.