Tumblr GIF Description Dataset

Overview

The Tumblr GIF (TGIF) dataset contains 100K animated GIFs and 120K sentences describing visual content of the animated GIFs. The animated GIFs have been collected from Tumblr, from randomly selected posts published between May and June of 2015. We provide the URLs of animated GIFs in this release. The sentences are collected via crowdsourcing, with a carefully designed annotation interface that ensures high quality dataset. We provide one sentence per animated GIF for the training and validation splits, and three sentences per GIF for the test split. The dataset shall be used to evaluate animated GIF/video description techniques.

Basic Statistics

Splits	Train	Validation	Test	Overall
# Animated GIFs	80,000	10,708	11,360	102,068
# Frames (Total/Average)	3,258,373/40.76	431,359/40.30	453,718/39.96	4,143,450/40.62
# Shots (Total/Average)	204,553/2.56	26,559/2.48	27,933/2.46	259,338/2.54
Duration (Total/Average)	81h/3.66s	11h/3.60s	12h/3.65s	103h/3.65s
# Sentences (Total/Average)	80,000/1	10,708/1	34,101/3	125,782/1.23
# Tokens (Total/Median)	911,593/11	10,831/11	34,101/11	1,418,555/11
# Unique Tokens	10,685	4,755	7,083	12,228
Average Term Frequency	85.32	25.72	54.31	116.00

Download

Research Paper Dataset Source code Pretrained model Evaluation suite Crowdsourcing pipeline Animated GIF filtering

Citation

@InProceedings{tgif-cvpr2016,
  author = {Li, Yuncheng and Song, Yale and Cao, Liangliang and Tetreault, Joel and Goldberg, Larry and Jaimes, Alejandro and Luo, Jiebo},
  title = "{TGIF: A New Dataset and Benchmark on Animated GIF Description}",
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2016}
}