metadata

annotations_creators:
  - crowdsourced
language:
  - en
language_creators:
  - found
license:
  - cc-by-4.0
multilinguality:
  - monolingual
paperswithcode_id: vasr
pretty_name: VASR
size_categories:
  - 1K<n<10K
source_datasets:
  - original
tags:
  - commonsense-reasoning
  - visual-reasoning
task_ids: []
extra_gated_prompt: >-
  By clicking on “Access repository” below, you also agree that you are using it
  solely for research purposes. The full license agreement is available in the
  dataset files.

Dataset Card for VASR

Dataset Description
Dataset Structure
- Data Fields
- Data Splits
Dataset Creation
Considerations for Using the Data
- Licensing Information
- Citation Information

Dataset Description

VASR is a challenging dataset for evaluating computer vision commonsense reasoning abilities. Given a triplet of images, the task is to select an image candidate B' that completes the analogy (A to A' is like B to what?). Unlike previous work on visual analogy that focused on simple image transformations, we tackle complex analogies requiring understanding of scenes. Our experiments demonstrate that state-of-the-art models struggle with carefully chosen distractors (±53%, compared to 90% human accuracy).

Homepage: https://vasr-dataset.github.io/
Colab https://colab.research.google.com/drive/1HUg0aHonFDK3hVFrIRYdSEfpUJeY-4dI
Repository: https://github.com/vasr-dataset/vasr/tree/main/experiments
Paper: NA
Leaderboard: https://vasr-dataset.github.io/
Point of Contact: yonatanbitton1@gmail.com

Supported Tasks and Leaderboards

https://vasr.github.io/leaderboard. https://paperswithcode.com/dataset/vasr.

Colab notebook code for VASR evaluation with ViT

https://colab.research.google.com/drive/1HUg0aHonFDK3hVFrIRYdSEfpUJeY-4dI

Languages

English.

Dataset Structure

Data Fields

A: datasets.Image() - the first input image, A:A'
A': datasets.Image() - the second input image, different from A' in a single key, A:A'
B: datasets.Image() - the third input image, has the same different item as A, B:B'
B': datasets.Image() - the forth image, which is the analogy solution. Different from B' in a single key (the same different one as in A:A'), B:B'
candidates_images: [datasets.Image()] - a list of candidate images solutions to the analogy
label: datasets.Value("int64") - the index of the ground-truth solution
candidates: [datasets.Value("string")] - a list of candidate string solutions to the analogy
A_verb: datasets.Value("string") - the verb of the first input image A
A'_verb: datasets.Value("string") - the verb of the second input image A'
B_verb: datasets.Value("string") - the verb of the third input image B
B'_verb: datasets.Value("string") - the verb of the forth image, which is the analogy solution
diff_item_A: datasets.Value("string") - FrameNet key of the item that is different between A:A', in image A (which is the same as image B)
diff_item_A_str_first: datasets.Value("string") - String representation of the FrameNet key of the item that is different between A:A', in image A
diff_item_A': datasets.Value("string") - FrameNet key of the item that is different between A:A', in image A' (which is the same as image B')
diff_item_A'_str_first: datasets.Value("string") - String representation of the FrameNet key of the item that is different between A:A', in image A'

Data Splits

There are three splits, TRAIN, VALIDATION, and TEST.
Since there are four candidates and one solution, random chance is 25%.

Dataset Creation

We leverage situation recognition annotations and the CLIP model to generate a large set of 500k candidate analogies. There are two types of labels:

Silver labels, obtained from the automatic generation.
Gold labels, obtained from human annotations over the silver annotations.

In the huggingface version we provide only the gold labeled dataset. Please refer to the project website download page if you want to download the silver labels version.

Annotations

Annotation process

We paid Amazon Mechanical Turk Workers to solve analogies, five annotators for each analogy. Workers were asked to select the image that best solves the analogy. The resulting dataset is composed of the 3,820 instances agreed upon with a majority vote of at least 3 annotators, which was obtained in 93% of the cases.