Does knowing the question help in image understanding?

Dataset

MS COCO VQA dataset

Format

Image followed by its three questions-answers pairs.

Example

  • Question: What is the mustache made of?
  • Inference: A male is present who has a mustache
  • Truth: Girl is wearing banana mustache
  • Other question: What color are her eyes? (Female with mustache can be inferred from this)

  • Question: Do you like this house?
  • Inference: Outdoor scene showing a house.
  • Truth: No help while answering other questions.
  • Other questions: Who is riding a bike? Is the boy on a skateboard?

Observations

  • If the questions are independent then they don’t help each other but maybe they can help in image understanding
  • Make a BOW model with the questions and then derive inferences to make one single description. This description can be appended to the image features.
  • Use Stanford CoreNLP to convert question to description, this description might give some more information about image
  • Use a captioning system to generate some assumed truth about the system and compare this with the questions. This will generate a different feature set.
  • If we use a BOW model, then we will need a vocabulary consisting of all possible words. Maybe a word embedding approach will work here.
  • How to use the given question categories (64)

Help Score

Dataset is shuffled while maintaining order of image id with its respective question and label. The labels are obtained by manual annotation following rules described. All questions are provided to the user.

Question - Image Help

If presence of any question helps in understanding the image and the objects present in the image then this set of questions is not independent, so is given a help score of 1. The help score is defined as one, if reading all the questions gives at least some information about the image. Manually checking all the image-question sets we get a help score of 122 out of 200 images. So, out of 200 images there were 122 images whose questions provided some information to understand the image. Example information can be understanding if the scene is outdoor or indoor or anything else. A help score of 0 was given to ambiguous questions. It is true that images with help score of one may be given wrong hint from some question, however we are only checking if any kind of help is provided.

Help Score = 122 / 200

Question - Question Help

If presence of any question helps in answering other questions, then this set is not independent, so is given a help score of 1. The help score is defined as one, if reading all the questions can help in answering at least one of the questions. Manually checking all the question sets, we get a help score of 109 out of 200 images. So, out of 200 images there were 109 images whose questions provided some information useful for other questions. A help score of 0 was given to ambiguous questions.

Help Score = 109 / 200

TODO

5/23
  • [X] Run NLP baseline using Word2Vec
  • [ ] CV baseline
5/18
  • [X] Image-Questions Help Score Annotations


blog comments powered by Disqus