论文网址：Visual Questions Answering Developments, Applications, Datasets and Opportunities: A State-of-the-Art Survey | IEEE Conference Publication | IEEE Xplore

英文是纯手打的！论文原文的summarizing and paraphrasing。可能会出现难以避免的拼写错误和语法错误，若有发现欢迎评论指正！文章偏向于笔记，谨慎食用

目录

1. 心得

2. 论文逐段精读

2.1. Abstract

2.2. Introduction

2.3. Traditional Techniques For Visual Question Answering

2.4. Recent Developments In Visual Question Answering

2.4.1. General VQA Model

2.4.2. Embedding Layer

2.4.3. Recurrent Neural Network

2.4.4. Randomly initialized Embedding layer + LSTM

2.5. Applications of Visual Question Answering

2.5.1. Medical Visual Question Answering(VQA) Model

2.5.2. VQA in Advertisement

2.6. Conclusion and Future Work

3. Reference

1. 心得

（1）是人类写的论文，是AI写不出来的离奇脑回路程度

2. 论文逐段精读

2.1. Abstract

        ①emm非常非常泛泛地说了一下对哪个方向进行调查回顾了一些进展总结了现存问题。说了和没说差不多

2.2. Introduction

        ①Main categories of VQA: 

Identifying	e.g. the objects in an image.
Detecting objects	e.g. Is there a car in the picture?
Classification of attributes	e.g. What color is the car?
Scene categorization	e.g. Is it raining?
Counting	e.g. How many cars are seen in the picture?

or other multi-hop inference needed questions

        ②VQA, an example that consists of CNN and RNN:

2.3. Traditional Techniques For Visual Question Answering

        ①说了一大坨话但没什么重要的东西

        ②Model performance on VQA dataset:

（什么玩意儿啊这是一个是数据表一个是柱状图？但是数据都一样？我大概比对了一下好像是一样的）

2.4. Recent Developments In Visual Question Answering

        ①啊好烦啊能不能别写这么大段话看着ex一口气读下来差点没给我闷死我都不知道在哪里可以喘气怎么可以把一堆数据集和一堆方法放在同一个段落里面的

2.4.1. General VQA Model

        ①不能引用一下自己的图片吗fig3全文就没有提到过：

图片裁剪和每个像素点减去平均值然后送进ResNet

2.4.2. Embedding Layer

        ①把问题那一句话转为嵌入

2.4.3. Recurrent Neural Network

        ①用预训练的RNN来生成预测：

2.4.4. Randomly initialized Embedding layer + LSTM

        ①？？？什么玩意2023了还在LSTM玩考古的baby吗

2.5. Applications of Visual Question Answering

2.5.1. Medical Visual Question Answering(VQA) Model

        ①图7？？？哥们儿连数字都数不清楚了？是典型的VQA技术：

作者提出的这个DAE可以无监督地使用很多医学无标签数据（怎么能突然转到医学的）

2.5.2. VQA in Advertisement

        ①举例了一些网络说了图8但根本就没有这张图

2.6. Conclusion and Future Work

        ①把上一章总结的局限放在这里好了：Limited Generalization, Ambiguity, Biases, Limited Understanding of Images, Limited Understanding of Language, Lack of Diverse Datasets

3. Reference

@INPROCEEDINGS{10104870,
  author={Singh, Harsimran Jit and Bathla, Gourav and Mehta, Munish and Chhabra, Gunjan and Singh, Pardeep},
  booktitle={2023 International Conference on Sustainable Computing and Data Communication Systems (ICSCDS)}, 
  title={Visual Questions Answering Developments, Applications, Datasets and Opportunities: A State-of-the-Art Survey}, 
  year={2023},
  volume={},
  number={},
  pages={778-785},
  keywords={Measurement;Visualization;Art;Training data;Machine learning;Question answering (information retrieval);Data communication;Visual Question Answering;Machine Learning;Deep Learning;Natural Language Processing},
  doi={10.1109/ICSCDS56580.2023.10104870}}