Awesome-Attention-Mechanism-in-cv

Table of Contents

• Introduction
• Attention Mechanism
• Plug and Play Module
• Evaluation
• Paper List
• Contribute

Introduction

PyTorch实现多种计算机视觉中网络设计中用到的Attention机制，还收集了一些即插即用模块。由于能力有限精力有限，可能很多模块并没有包括进来，有任何的建议或者改进，可以提交issue或者进行PR。

Attention Mechanism

Paper	Publish	Link	Main Idea	Blog
Global Second-order Pooling Convolutional Networks	CVPR19	GSoPNet	将高阶和注意力机制在网络中部地方结合起来
Neural Architecture Search for Lightweight Non-Local Networks	CVPR20	AutoNL	NAS+LightNL
Squeeze and Excitation Network	CVPR18	SENet	最经典的通道注意力	zhihu
Selective Kernel Network	CVPR19	SKNet	SE+动态选择	zhihu
Convolutional Block Attention Module	ECCV18	CBAM	串联空间+通道注意力	zhihu
BottleNeck Attention Module	BMVC18	BAM	并联空间+通道注意力	zhihu
Concurrent Spatial and Channel ‘Squeeze & Excitation’ in Fully Convolutional Networks	MICCAI18	scSE	并联空间+通道注意力	zhihu
Non-local Neural Networks	CVPR19	Non-Local(NL)	self-attention	zhihu
GCNet: Non-local Networks Meet Squeeze-Excitation Networks and Beyond	ICCVW19	GCNet	对NL进行改进	zhihu
CCNet: Criss-Cross Attention for Semantic Segmentation	ICCV19	CCNet	对NL改进
SA-Net:shuffle attention for deep convolutional neural networks	ICASSP 21	SANet	SGE+channel shuffle	zhihu
ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks	CVPR20	ECANet	SE的改进
Spatial Group-wise Enhance: Improving Semantic Feature Learning in Convolutional Networks	CoRR19	SGENet	Group+spatial+channel
FcaNet: Frequency Channel Attention Networks	CoRR20	FcaNet	频域上的SE操作
$A^2\text{-}Nets$: Double Attention Networks	NeurIPS18	DANet	NL的思想应用到空间和通道
Asymmetric Non-local Neural Networks for Semantic Segmentation	ICCV19	APNB	spp+NL
Efficient Attention: Attention with Linear Complexities	CoRR18	EfficientAttention	NL降低计算量
Image Restoration via Residual Non-local Attention Networks	ICLR19	RNAN
Exploring Self-attention for Image Recognition	CVPR20	SAN	理论性很强，实现起来很简单
An Empirical Study of Spatial Attention Mechanisms in Deep Networks	ICCV19	None	MSRA综述self-attention
Object-Contextual Representations for Semantic Segmentation	ECCV20	OCRNet	复杂的交互机制，效果确实好
IAUnet: Global Context-Aware Feature Learning for Person Re-Identification	TTNNLS20	IAUNet	引入时序信息
ResNeSt: Split-Attention Networks	CoRR20	ResNeSt	SK+ResNeXt
Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks	NeurIPS18	GENet	SE续作
Improving Convolutional Networks with Self-calibrated Convolutions	CVPR20	SCNet	自校正卷积
Rotate to Attend: Convolutional Triplet Attention Module	WACV21	TripletAttention	CHW两两互相融合
Dual Attention Network for Scene Segmentation	CVPR19	DANet	self-attention
Relation-Aware Global Attention for Person Re-identification	CVPR20	RGANet	用于reid
Attentional Feature Fusion	WACV21	AFF	特征融合的attention方法
An Attentive Survey of Attention Models	CoRR19	None	包括NLP/CV/推荐系统等方面的注意力机制
Stand-Alone Self-Attention in Vision Models	NeurIPS19	FullAttention	全部的卷积都替换为self-attention
BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmentation	ECCV18	BiSeNet	类似FPN的特征融合方法	zhihu
DCANet: Learning Connected Attentions for Convolutional Neural Networks	CoRR20	DCANet	增强attention之间信息流动
An Empirical Study of Spatial Attention Mechanisms in Deep Networks	ICCV19	None	对空间注意力进行针对性分析
Look closer to see better: Recurrent attention convolutional neural network for fine-grained image recognition	CVPR17 Oral	RA-CNN	细粒度识别
Guided Attention Network for Object Detection and Counting on Drones	ACM MM20	GANet	处理目标检测问题
Attention Augmented Convolutional Networks	ICCV19	AANet	多头+引入额外特征映射
GLOBAL SELF-ATTENTION NETWORKS FOR IMAGE RECOGNITION	ICLR21	GSA	新的全局注意力模块
Attention-Guided Hierarchical Structure Aggregation for Image Matting	CVPR20	HAttMatting	抠图方面的应用，高层使用通道注意力机制，然后再使用空间注意力机制指导低层。
Weight Excitation: Built-in Attention Mechanisms in Convolutional Neural Networks	ECCV20	None	与SE互补的权值激活机制
Expectation-Maximization Attention Networks for Semantic Segmentation	ICCV19 Oral	EMANet	EM+Attention
Dense-and-implicit attention network	AAAI 20	DIANet	LSTM+全程SE注意力
Coordinate Attention for Efficient Mobile Network Design	CVPR21	CoordAttention	横向、竖向
Cross-channel Communication Networks	NIPS19	C3Net	GNN+SE
Gated Convolutional Networks with Hybrid Connectivity for Image Classification	AAAI20	HCGNet	引入了LSTM的部分概念
Weighted Channel Dropout for Regularization of Deep Convolutional Neural Network	AAAI19	None	Dropout+SE
BA^2M: A Batch Aware Attention Module for Image Classification	CVPR21	None	Batch之间建立attention
EPSANet：An Efficient Pyramid Split Attention Block on Convolutional Neural Network	CoRR21	EPSANet	多尺度
Stand-Alone Self-Attention in Vision Models	NIPS19	SASA	Non-Local变体
ResT: An Efficient Transformer for Visual Recognition	CoRR21	ResT	self-attention变体
Spanet: Spatial Pyramid Attention Network for Enhanced Image Recognition	ICME20	SPANet	多个AAP组成金字塔
Space-time Mixing Attention for Video Transformer	CoRR21	X-VIT Not release	VIT+时空attention
DMSANet: Dual Multi Scale Attention Network	CoRR21	Not release yet	两尺度+轻量
CompConv: A Compact Convolution Module for Efficient Feature Learning	CoRR21	Not release yet	res2net+ghostnet
VOLO: Vision Outlooker for Visual Recognition	CoRR21	VOLO	ViT上的Attention
Interflow: Aggregating Multi-layer Featrue Mappings with Attention Mechanism	CoRR21	Not release yet	辅助头级别attention
MUSE: Parallel Multi-Scale Attention for Sequence to Sequence Learning	CoRR21	MUSE Attention	NLP中对SA进行改进
Polarized Self-Attention: Towards High-quality Pixel-wise Regression	CoRR21	PSA	Pixel-wise regression
CA-Net: Comprehensive Attention Convolutional Neural Networks for Explainable Medical Image Segmentation	TMI21	CA-Net	Spatial Attention
BAM: A Lightweight and Efficient Balanced Attention Mechanism for Single Image Super Resolution	CoRR21	BAM	Super resolution
Attention as Activation	CoRR21	ATAC	activation + attention
Region-based Non-local Operation for Video Classification	CoRR21	RNL	video classification
MSAF: Multimodal Split Attention Fusion	CoRR21	MSAF	MultiModal
All-Attention Layer	CoRR19	None	Tranformer Layer
Compact Global Descriptor	CoRR20	CGD	add every two channel attention
SimAM: A Simple, Parameter-Free Attention Module for Convolutional Neural Networks	ICML21	SimAM	类脑计算神经元能量
Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks With Octave Convolution	ICCV19	OctConv	从频率角度改进
Contextual Transformer Networks for Visual Recognition	ICCV2021	CoTNet	虽然宣称Transformer改进，但实际上就是non-local非常接近
Residual Attention: A Simple but Effective Method for Multi-Label Recognition	ICCV2021	CSRA	用于多标签图像识别任务
Self-supervised Equivariant Attention Mechanism for Weakly Supervised Semantic Segmentation	CVPR2020	SEAM	弱监督
An Attention Module for Convolutional Neural Networks	ICCV2021	AW-Conv	提升了SE部分的容量

Plug and Play Module

• ACBlock
• Swish、wish Activation
• ASPP Block
• DepthWise Convolution
• Fused Conv & BN
• MixedDepthwise Convolution
• PSP Module
• RFBModule
• SematicEmbbedBlock
• SSH Context Module
• Some other usefull tools such as concate feature map、flatten feature map
• WeightedFeatureFusion:EfficientDet中的FPN用到的fuse方式
• StripPooling：CVPR2020中核心代码StripPooling
• GhostModule: CVPR2020GhostNet的核心模块
• SlimConv: SlimConv3x3
• Context Gating： video classification
• EffNetBlock: EffNet
• ECCV2020 BorderDet: Border aligment module
• CVPR2019 DANet: Dual Attention
• Object Contextual Representation for sematic segmentation: OCRModule
• FPT: 包含Self Transform、Grounding Transform、Rendering Transform
• DOConv: 阿里提出的Depthwise Over-parameterized Convolution
• PyConv: 起源人工智能研究院提出的金字塔卷积
• ULSAM：用于紧凑型CNN的超轻量级子空间注意力模块
• DGC: ECCV 2020用于加速卷积神经网络的动态分组卷积
• DCANet: ECCV 2020 学习卷积神经网络的连接注意力
• PSConv: ECCV 2020 将特征金字塔压缩到紧凑的多尺度卷积层中
• Dynamic Convolution: CVPR2020 动态滤波器卷积（非官方）
• CondConv: Conditionally Parameterized Convolutions for Efficient Inference

Evaluation

基于CIFAR10+ResNet+待测评模块，对模块进行初步测评。测评代码来自于另外一个库：https://github.com/kuangliu/pytorch-cifar/ 实验过程中，不使用预训练权重，进行随机初始化。

模型	top1 acc	time	params(MB)
SENet18	95.28%	1:27:50	11,260,354
ResNet18	95.16%	1:13:03	11,173,962
ResNet50	95.50%	4:24:38	23,520,842
ShuffleNetV2	91.90%	1:02:50	1,263,854
GoogLeNet	91.90%	1:02:50	6,166,250
MobileNetV2	92.66%	2:04:57	2,296,922
SA-ResNet50	89.83%	2:10:07	23,528,758
SA-ResNet18	95.07%	1:39:38	11,171,394

Contribute

欢迎在issue中提出补充的文章paper和对应code链接。