VisKE: Visual Knowledge Extraction and Question Answering by Visual Verification of Relation Phrases (Relation Phrases)
Created by Fereshteh Sadeghi at University of Washington.
This repository contains the list of relation phrases used in the VisKE paper along with the confidence scores.
Relation phrases in this set are gathered using the Google Books Ngram (English 2012).
Relation phrases in this set are generated by randomly permuting different subjects with different verbs and objects.
Each relation phrase consists of a Subject, Verb and Object. Each relation has a label which shows if the phrase is a correct statement (1) or not (-1). For each phrase we have provided the scores obtained from MSR Language Model (MSRLM) as a baseline as well as scores obtained from our proposed method (VisKE). For each Subject (animal) the relation phrases are gathered in a separate csv file with the following format:
Subject, Verb, Object, Label, MSRLM Score, VisKE Score
Function computePR is used to compute the average precision.
This is a post-CVPR cleaned-up version of the data and may produce slightly different results than the performance reported in the paper.
Please cite this paper if you use the data in your reseach. http:https://homes.cs.washington.edu/~fsadeghi/papers/fsadeghi_VisKE.pdf
@inproceedings{sadeghi2015viske,
title={VisKE: Visual Knowledge Extraction and Question Answering by Visual Verification of Relation Phrases},
author={Sadeghi, Fereshteh and Divvala, Santosh K and Farhadi, Ali},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
pages={1456--1464},
year={2015}
}