论文网址：pdf

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. Background and Motivation

2.3.1. Motivation

2.4. CLIP-ViL

2.4.1. Visual Question Aswering

2.4.2. Image Captioning

2.4.3. Vision-and-Language Navigation

2.5. Vision-and-Language Pre-training

2.5.1. CLIP-VIL_p

2.5.2. Experiments

2.6. Analysis

2.7. Conclusions

1. 心得

（1）？非常简单的一篇文章，感觉在测试CLIP？

2. 论文逐段精读

2.1. Abstract

        ①Model pre-trained on large number of data brings better performance

        ②Scenarios suitable for CLIP: plug and fine-tune, or combining with V&L

2.2. Introduction

        ①Bottleneck of vision-and-language (V&L) tasks: visual representation and scarce labled data

        ②Most V&L tasks require complex reasoning, which can not use visual model directly

        ③They define two scenarios:

CLIP_ViL	CLIP in direct task-specific fine-tuning
CLIP_ViL_p	integrate CLIP with V&L pre-training on image-text pairs and transfer to downstream tasks

        ④Tasks: Visual Question Answering, Image Captioning, and Vision-and-Language Navigation

2.3. Background and Motivation

        ①Training stage: 

visual encoder pretrianing, alignment (opt), downstream task

        ②Different types of model:

region based, network based, and CLIP (contrastive)

2.3.1. Motivation

        ①就是说直接把CLIP用在不同复杂视觉任务上性能一般般所以要小改一下

2.4. CLIP-ViL

2.4.1. Visual Question Aswering

        ①Performance of models on VQA v2.0 dataset:

2.4.2. Image Captioning

        ①Image captioning comparison table on COCO dataset:

2.4.3. Vision-and-Language Navigation

        ①The model performance on Room-to-Room (R2R) dataset:

        ②Changing ResNet to CLIP, the performance table:

2.5. Vision-and-Language Pre-training

2.5.1. CLIP-VIL_p

        ①For text segment , tokenize it into subwords  and further embedded as the sum of its token, position and segment embeddings 

        ②Image  is is embedded as 

        ③Concatenate them two as 

        ④Reconstruct sentence with 15% mask ratio, match text and image with the 50% correct sentence ratio, then execute visual question answering

2.5.2. Experiments

        ①Two variants of CLIP as visual encoder: CLIP-Res50andCLIP Res50x4

        ②Datasets: MSCOCOCaptions, VisualGenomeCaptions, VQA,GQA, and VG-QA  for pre-training

        ③Patch number for each image: 100

        ④Epoch of pretraining: 20

        ⑤Fine tune pretrained model on evaluation stage

        ⑥Dataset of tasks: VQAv2.0, visual entailment SNLI-VE, and GQA

        ⑦Results:

2.6. Analysis

        ①Zero-shot performance of CLIP on VQA v2.0 mini-eval:

        ②Influence of V&L pre-training:

        ③Visualization of feature positioning of different models:

2.7. Conclusions

        ~