Text Encoder

使用的backbone是BERT(通过MLM训练)
该研究认为，image encoder的模型大小应该大于text encoder,所以在text encoder这里，只使用六层self attention来提取特征，剩余六层cross attention用于multi-modal encoder。

ITC Loss & Momentum

参考Moco [[Moco- Momentum Contrast for Unsupervised Visual Representation Learning]]

Improve Noisy Web Data

见[[BLIP]]，是沿用的工作

Loss

ITM

MLM

A vision-language model that unifies vision-language understanding and generation tasks.

主要分为两块工作：

• 去除图文检索所使用的数据集中的噪声
• vision language understanding and generation

Model

Noise Filtering

Understanding & Generation

GLIP是一个学习了object-level, language-aware, and semantic-rich visual representations 的模型。
统一对象检测和短语接地进行预训练。

重要的问题

什么是 phrase grounding:
Phrase Grounding refers to the task of associating or “grounding” a natural language phrase (like a sentence or a word) to a specific region or object in an image. In other words, it’s about finding which part of the image corresponds to the object or concept described by a given text phrase.

For instance, if you have the phrase “the red ball on the table” and an image of a room with a red ball placed on a table, the goal of phrase grounding is to identify the exact region in the image that corresponds to the “red ball on the table”, distinguishing it from other objects in the image.

Unified Formulation

传统的物体检测方法会把每个region分类进c个classes，而本文使用的Object detection as phrase grounding.
我们通过将每个区域与文本提示中的c(class)短语进行接地/对齐，将检测重新制定为基础任务
the classification prompt “person. bicycle. car. … . toothbrush”

https://blog.csdn.net/h661975/article/details/135116957

loss: ITC (Image Text Contrastive)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

# image_encoder - ResNet or Vision Transformer 
# text_encoder - CBOW or Text Transformer 
# I[n, h, w, c] - minibatch of aligned images 
# T[n, l] - minibatch of aligned texts 
# W_i[d_i, d_e] - learned proj of image to embed 
# W_t[d_t, d_e] - learned proj of text to embed 
# t - learned temperature parameter  

# extract feature representations of each modality 
I_f = image_encoder(I) #[n, d_i] 
T_f = text_encoder(T) #[n, d_t]  

# joint multimodal embedding [n, d_e] 
I_e = l2_normalize(np.dot(I_f, W_i), axis=1) T
_e = l2_normalize(np.dot(T_f, W_t), axis=1)  

# scaled pairwise cosine similarities [n, n] 
logits = np.dot(I_e, T_e.T) * np.exp(t)  

# symmetric loss function 
labels = np.arange(n) 
loss_i = cross_entropy_loss(logits, labels, axis=0) 
loss_t = cross_entropy_loss(logits, labels, axis=1) 
loss = (loss_i + loss_t)/2

Cross_entropy_loss:

CLIP 本质上是全局图像嵌入，不利于像素对齐特征提取。