Memento10k

Memorability dataset with 10000 3-second videos. Each video has upwards of 90 human annotations, and the split-half consistency of this dataset is 0.73 (best in class for video memorabilty datasets).