Baseline: Show, Ask, Attend, and Answer: A Strong Baseline For Visual Question Answering (paper: https://arxiv.org/pdf/1704.03162.pdf) (code: https://github.com/Cyanogenoid/pytorch-vqa)
Datasets: VQA2.0 http://www.visualqa.org/download.html
Evaluation: https://github.com/GT-Vision-Lab/VQA
Experiments To Do:
To run today:
-
Add converse substitution to Language Only augmentation
-
Multiple word substitutions.
-
Do some paraphrasing for known question types.
-
How many / Color of - question substitution with hypernym doubt.
-
Language augment other methods.
-
Change all augmentation methods to fit the same vocab.
-
Filter conceptnet based on question repetition.
-
Add all working methods together for data augmentation.
-
Make custom test set for places 365? Places 365 has adjectives as well as scene understanding.
-
Add augmentation on image based on wrong answer or image type?