TVSum stands for Title-based Video Summarization Dataset, which is a benchmark for verifying video summarization technology.

The dataset contains 50 videos of different genres (e.g., news, how-to, documentary, vlog, selfie), and 1,000 crowdsourced annotations (20 per video) of the importance of the shots. The videos and annotation data allow for the automatic evaluation of various video summarization techniques without the need for (expensive) user studies.