Visual Commonsense Reasoning

On VCR, a model must not only answer commonsense visual questions, but also provide a rationale that explains why the answer is true.


Submitting to the leaderboard

Submission is easy! You just need to email Rowan with your predictions. Formatting instructions are below:

Please include in your email 1) a name for your model, 2) your team name (including your affiliation), and optionally, 3) a github repo or paper link.

I'll try to get back to you within a few days, usually sooner. Teams can only submit results from a model once every 7 days.

I reserve the right to not score any of your submissions if you cheat -- for instance, please don't make up a bunch of fake names / email addresses and send me multiple submissions under those names.


What kinds of submissions are allowed?

The only constraint is that your system must predict the answer first, then the rationale. (The rationales were selected to be highly relevant to the correct Q,A pair, so they leak information about the correct answer.)

  • To deter this, the submission format involves submitting predictions for each possible rationale, conditioned on each possible answer.
  • A simple way of setting up the experiments (used in the paper) is to consider a task with query and four response choices. For Q->A the query is the question, and the response choices are the answers. For QA->R, the query is the question and answer, concatenated together, and the response choices are the rationales.

Questions?

If it's not about something private, check out the google group below:

VCR Leaderboard

There are two different subtasks to VCR:

  • Question Answering (Q->A): In this setup, a model is provided a question, and has to pick the best answer out of four choices. Only one of the four is correct.
  • Answer Justification (QA->R): In this setup, a model is provided a question, along with the correct answer, and it has to justify it by picking the best rationale out of four choices.

We combine the two parts with the Q->AR metric, in which a model only gets a question right if it answers correctly and picks the right rationale. Models are evaluated in terms of accuracy (%). How well will your model do?

Rank Model Q->A QA->R Q->AR
Human Performance

University of Washington

(Zellers et al. '18)
91.0 93.0 85.0

📼

May 13, 2019
B2T2 (ensemble)

CJD

74.077.157.1

2

May 13, 2019
B2T2 (single model)

CJD

72.675.755.0

3

May 22, 2019
TNet (ensemble)

FlyingPig (Sun Yat-sen University)

72.772.653.0

4

May 22, 2019
B-VCR

DUAL

70.571.550.8

5

May 22, 2019
TNet (single model)

FlyingPig (Sun Yat-sen University)

70.970.650.4

6

May 13, 2019
HGL

HCP

70.170.849.8

7

May 18, 2019
CCD

Anonymous

68.570.548.4

8

May 16, 2019
MRCNet

MILAB (Seoul National University)

68.470.548.4

9

May 17, 2019
MUGRN

NeurIPS 2019 submission ID 21

68.269.447.5

10

May 13, 2019
SGRE

UTS

https://github.com/AmingWu/Multi-modal-Circulant-Fusion/
67.569.746.9

11

Feb 19, 2019
FAIR

Facebook AI Research

65.770.146.3

12

Feb 25, 2019
CKRE

Peking University

66.968.245.9

13

May 14, 2019
emnet

DCP

66.668.045.4

14

Nov 28, 2018
Recognition to Cognition Networks

University of Washington

https://github.com/rowanz/r2c
65.167.344.0

15

May 20, 2019
DVD

SL

66.365.043.3

16

March 27, 2019
GS Reasoning

UC San Diego

65.761.041.1

17

May 17, 2019
R2R (text only)

Anonymous

58.469.140.5

18

Nov 28, 2018
BERT-Base

Google AI Language (experiment by Rowan)

https://github.com/google-research/bert
53.964.535.0

19

Nov 28, 2018
MLB

Seoul National University (experiment by Rowan)

https://github.com/jnhwkim/MulLowBiVQA
46.236.817.2
Random Performance 25.0 25.0 6.2