Please cite the following papers when referring to the VidLife in academic publications and papers. VidLife includes 29 sorts of high-level activities and 14 sorts of low-level activities. ] 'įor the use of video clips with bounding box, subtitle and ground truth question-answer pair, please find TVQA+ dataset website. ' 'labeled_SVO_ts': ', ] ' 'labeled_train_SVO_ts': ', ] ' 'ground_truth_QA_pair(or sentence)': '. Data FormatĮach example in VidLife is a video clip with its corresponding bounding box, subtitle from TVQA+ dataset and multi-label life event annotation. In the paper, we initiate a novel life event extraction task and present a life event extraction system, LifeEventMiner as a baseline model. High-level activities which consist of several low-level activities express relatively complex life events. We distinguish life events into high-level life events and low-level life events, and regard an activity that comprises a specific action only, and the action acts on a concrete object as a low-level activity.

VidLife contains 14,343 training examples, 1,793 examples for validation and 1,793 examples for testing. The complete dataset will be made available online after our paper is accepted. You could download part of the life events annotations, which is released in this repository. (Leonard, visit, Penny), which is designed for training and evaluating personal life event extraction systems. VidLife contains personal life events with triple forms from The Big Bang Theory, eg.

Download project lifelog