Watch-n-Patch Action Recognition Dataset

The Watch-n-Patch dataset focuses on modeling human activities, covering a variety of actions in completely unsupervised situations. The data is collected by the Microsoft Kinect One sensor, with a total length of about 230 minutes and 458 videos. The dataset requires 7 subjects to perform daily activities in 8 offices and 5 kitchens with complex backgrounds. In addition, the dataset also provides skeleton data as ground truth annotations.