We randomly selected three videos from the Internet, that are longer than 1.5K frames and have their main objects continuously appearing. Each video has 20 uniformly sampled frames manually annotated for evaluation. Each video has been played back and forth to generate videos that are three times as long.