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

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”

,

通过结合[[DINO]]和grounded-pretraining，可以使用人类输入（例如类别名称或转介表达式）检测任意对象
Open-Vocab. Det

an open-set object detector that can detect any objects with respect to an arbitrary free-form text prompt. The model was trained on over 10 million images, including detection data, visual grounding data, and image-text pairs. It has a strong zero-shot detection performance. However, the model needs text as inputs and can only detect boxes with corresponding phrases.

Grounding-DINO

Principle

Tight modality fusion based on [[DINO]]

什么是feature fusion?

伪代码：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37

# f_q, f_k: encoder networks for query and key 
# queue: dictionary as a queue of K keys (CxK) 
# m: momentum 
# t: temperature  

f_k.params = f_q.params # initialize 
for x in loader: # load a minibatch x with N samples 
	x_q = aug(x) # a randomly augmented version 
	x_k = aug(x) # another randomly augmented version  
	q = f_q.forward(x_q) # queries: NxC 
	k = f_k.forward(x_k) # keys: NxC 
	k = k.detach() # no gradient to keys  
	
	# positive logits: Nx1 
	l_pos = bmm(q.view(N,1,C), k.view(N,C,1))  # 相当于把batch中每个正样本对之间求了cosine临近
	
	# negative logits: NxK 
	l_neg = mm(q.view(N,C), queue.view(C,K))  
	
	# logits: Nx(1+K) 
	logits = cat([l_pos, l_neg], dim=1)  
	
	# contrastive loss, Eqn.(1) 
		labels = zeros(N) # positives are the 0-th，将识别的类别视为0,可以直接使用CrossEntropyLoss
	loss = CrossEntropyLoss(logits/t, labels)  
	
	# SGD update: query network 
	loss.backward() 
	update(f_q.params)  
	
	# momentum update: key network 
	f_k.params = m*f_k.params+(1-m)*f_q.params  
	
	# update dictionary 
	enqueue(queue, k) # enqueue the current minibatch 
	dequeue(queue) # dequeue the earliest minibatch

亮点

Dictionary as a queue

在使用key encoder(momentum encoder)创建负样本，并把encode过的负样本存在一个queue（FIFO）中方便后续对比时直接使用，每次训练都会使用一个新的mini batch，此时会将此mini batch中的样本encode之后加入queue并删除存在最久的那个mini batch的样本（因为考虑到最老的mini batch使用的encoder是最过时的，所以FIFO是非常合理的），这样可以有效控制负样本的数量，也就是公式中的K。

• 节省字典的计算开销
• 而且mini batch大小可以直接和负样本脱钩

Momentum update

因为负样本数量（字典/队列）很大，所以没办法给key encoder回传梯度，所以可以考虑把query encoder的参数直接复制给key encoder，但过快改变的key encoder会导致样本字典的特征不一致，所以使用动量更新的方式。

过往工作对比

b)
只有一个编码器进行学习。Memory bank存下了所有样本的key。每当梯度回传后，会把memory bank被本次训练中被采样过的key使用新的encoder进行更新。

• 缺乏特帧一致性
• 需要训练一阵个epoch才能更新一遍memory bank

MoCo和memory bank 更接近，但是使用了queue dictionary和momentum update

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。