Visual Madlibs Q&A

Dataset


Version I:    madlibs_train_v1.zip

Version II:    madlibs_train_v2.zip

  • This version splits training data into 80% sub-training data and 20% validation data, in "Solving Visual Madlibs with Multiple Cues", BMVC 2016
  • The validation data consists of easy and hard multiple-choice question-answers. Users are encouraged to use them to cross-validate the hyper-parameters.

Python tool:   

  • Used to access our dataset.


Evaluation


Task 1:    task1.zip

Task 2:    (including easy and hard version)

  • Task 1 is automatic targeted descriptions of images to fill in the blank.
  • Task 2 is targeted multiple-choice question answering of images.
  • For both tasks, we provide the image, instruction, and a Madlibs prompt.
  • For some type of question, we also provide the indication of targeted person or object.


Paper


Original Dataset Paper:   paper

@ARTICLE{VisualMadlibs, 
author = {Licheng Yu and Eunbyung Park and Alexander C. Berg and Tamara L. Berg}, 
title = "{Visual Madlibs: Fill in the blank Image Generation and Question Answering}", 
journal = {ICCV}, 
year = {2015}, 
} 

The Version II is provided by:   paper

@INPROCEEDINGS{VisualMadlibs_bmvc16, 
author = {Tatiana Tommasi, Arun Mallya, Bryan Plummer, Svetlana Lazebnik, Alex Berg, Tamara Berg}, 
title = "{Solving Visual Madlibs with Multiple Cues}", 
booktitle = {BMVC}, 
year = {2016}, 
} 
@ARTICLE{VisualMadlibs_ijcvSubmitted, 
author = {Tatiana Tommasi, Arun Mallya, Bryan Plummer, Svetlana Lazebnik, Alex Berg, Tamara Ber}, 
title = "{Solving Visual Madlibs with Multiple Cues}", 
journal = {arXiv preprint arXiv:1611.00393}, 
year = {2016}, 
}