VQA数据集调研报告
2018年3月16日
14:22
一、VQA数据集
1.与问题相关的数据集(related to questions)
1.1.训练集(training set)
(1)基本情况介绍:
name:v2_OpenEnded_mscoco_train2014_questions.json
size:40M
version: 2.0
year: 2017
contributor: VQA Team
date by created: 2017-04-26 17:07:13
task type: Open-Ended
data type: mscoco
url: http://visualqa.org
(2)形式(eg):
Name |
Type 类型 |
Description 描述 |
image_id |
int |
图片ID |
question_id |
|
问题的ID |
question |
str |
图片对应的问题 |
{"image_id": 458752,"question": "What is this photo taken looking through?","question_id": 458752000},{"image_id": 458752, "question": "What position isthis man playing?", "question_id": 458752001},{"image_id": 458752, "question": "What color is theplayers shirt?", "question_id": 458752002},{"image_id": 458752, "question": "Is this man aprofessional baseball player?", "question_id": 458752003}
(3)每张图片有一个编号,每张图片对应有若干个不同的问题,每个问题有一个编号,问题编号在图片编号的基础上增加三位,依次编为xxxxxx000,xxxxxx001 ...等等。
1.2.验证集(val)
(1)基本情况:除size为19.3M外其他情况同上
(2)形式:同上
1.3.测试集(test)
(1)开发测试集:除size为9.57M外其他同上
(2)测试集:除size为39.8M外其他同上
eg:{"image_id": 1, "question":"What is the fence made of?", "question_id": 1000},{"image_id": 1, "question": "What color is thetruck?", "question_id": 1001}, {"image_