The UT-Interaction human interaction dataset contains videos of six types of human-to-human interactions (continuous actions): handshake, pointing, hugging, pushing, kicking, and punching. The dataset provides the ground truth labels of these interactive actions, including time intervals and bounding boxes.

UT-Interaction has a total of 20 video sequences of about 1 minute in length. Each video contains at least one interactive action execution, and each video contains an average of 8 action executions. Several participants appear in the video, with more than 15 types of clothing. The video uses a resolution of 720*480, 30fps, and the height of the characters in the video is about 200 pixels.