Receptive Field Block Net for Accurate and Fast Object Detection

简介

本文在SSD基础上提出了RFB Module,利用神经科学的先验知识来解释这种效果提升。本质上是设计一种新的结构来提升感受野,并表明了人类视网膜的感受野有一个特点,离视线中心越远,其感受野是越大的,越靠近视线中间,感受野越小。基于此,本文提出的RFB Module就是来模拟人类这种视觉特点的。

[论文理解] Receptive Field Block Net for Accurate and Fast Object Detection-LMLPHP

RFB Module

结构如下图所示。

[论文理解] Receptive Field Block Net for Accurate and Fast Object Detection-LMLPHP

为什么要用空洞卷积呢?

首先要提高感受野,直观的想法就是要么加深层数,要么使用更大的卷积核,要么就是卷积之前使用pooling。加深层数网络参数就会变多,没法完成轻量级的任务;更大的卷积核一样参数也会变多;pooling虽然不会增加参数,但是会使信息损失,不利于后面层的信息传递。所以作者这里很自然的想到用空洞卷积,既不增加参数量,又能够提高感受野。

为什么要用这种多分支结构呢?

这是为了捕捉不同感受野的信息,如前面提到的,人类视野的特点就是距视野中心距离不同感受野不同,所以使用多分支结构,每个分支捕捉一种感受野,最后通过concat来融合感受野信息,就能达到模拟人类视觉的效果了。作者这里也给了一张图来说明。

[论文理解] Receptive Field Block Net for Accurate and Fast Object Detection-LMLPHP

为什么要提出两种版本的RFB呢?

左边的结构是原始的RFB,右边的结构相比RFB把3×3的conv变成了两个1×3和3×1的分支,一是减少了参数量,二是增加了更小的感受野,这样也是在模拟人类视觉系统,捕捉更小的感受野。

网络结构

整体网络结构如下所示,很好理解。

[论文理解] Receptive Field Block Net for Accurate and Fast Object Detection-LMLPHP

前面就是vgg19,然后从中间的层分出6个预测分支,比较好理解没啥记的。

代码复现

import torch
import torch.nn as nn
import torch.nn.functional as F
from torchsummary import summary
class RFBModule(nn.Module):
def __init__(self,out,stride = 1):
super(RFBModule,self).__init__()
self.s1 = nn.Sequential(
nn.Conv2d(out,out,kernel_size = 1),
nn.Conv2d(out,out,kernel_size=3,dilation = 1,padding = 1,stride = stride)
)
self.s2 = nn.Sequential(
nn.Conv2d(out,out,kernel_size =1),
nn.Conv2d(out,out,kernel_size=3,padding = 1),
nn.Conv2d(out,out,kernel_size=3,dilation = 3,padding = 3,stride = stride)
)
self.s3 = nn.Sequential(
nn.Conv2d(out,out,kernel_size =1),
nn.Conv2d(out,out,kernel_size = 5,padding =2),
nn.Conv2d(out,out,kernel_size=3,dilation=5,padding = 5,stride = stride)
)
self.shortcut = nn.Conv2d(out,out,kernel_size = 1,stride = stride)
self.conv1x1 = nn.Conv2d(out*3,out,kernel_size =1)
def forward(self,x):
s1 = self.s1(x)
s2 = self.s2(x)
s3 = self.s3(x)
#print(s1.size(),s2.size(),s3.size())
mix = torch.cat([s1,s2,s3],dim = 1)
mix = self.conv1x1(mix)
shortcut = self.shortcut(x)
return mix + shortcut
class RFBsModule(nn.Module):
def __init__(self,out,stride = 1):
super(RFBsModule,self).__init__()
self.s1 = nn.Sequential(
nn.Conv2d(out,out,kernel_size = 1),
nn.Conv2d(out,out,kernel_size=3,dilation = 1,padding = 1,stride = stride)
)
self.s2 = nn.Sequential(
nn.Conv2d(out,out,kernel_size =1),
nn.Conv2d(out,out,kernel_size=(1,3),padding = (0,1)),
nn.Conv2d(out,out,kernel_size=3,dilation = 3,padding = 3,stride = stride)
)
self.s3 = nn.Sequential(
nn.Conv2d(out,out,kernel_size =1),
nn.Conv2d(out,out,kernel_size = (3,1),padding =(1,0)),
nn.Conv2d(out,out,kernel_size=3,dilation=3,padding = 3,stride = stride)
)
self.s4 = nn.Sequential(
nn.Conv2d(out,out,kernel_size =1),
nn.Conv2d(out,out,kernel_size=3),
nn.Conv2d(out,out,kernel_size = 3,dilation = 5,stride = stride,padding = 6)
)
self.shortcut = nn.Conv2d(out,out,kernel_size = 1,stride = stride)
self.conv1x1 = nn.Conv2d(out*4,out,kernel_size =1)
def forward(self,x):
s1 = self.s1(x)
s2 = self.s2(x)
s3 = self.s3(x)
s4 = self.s4(x)
#print(s1.size(),s2.size(),s3.size(),s4.size())
#print(s1.size(),s2.size(),s3.size())
mix = torch.cat([s1,s2,s3,s4],dim = 1)
mix = self.conv1x1(mix)
shortcut = self.shortcut(x)
return mix + shortcut class RFBNet(nn.Module):
def __init__(self):
super(RFBNet,self).__init__()
self.feature_1 = nn.Sequential(
nn.Conv2d(3,64,kernel_size = 3,padding = 1),
nn.ReLU(),
nn.Conv2d(64,64,kernel_size=3,padding=1),
nn.ReLU(),
nn.MaxPool2d(kernel_size = 2,stride = 2),
nn.Conv2d(64,128,kernel_size = 3,padding = 1),
nn.ReLU(),
nn.Conv2d(128,128,kernel_size=3,padding=1),
nn.ReLU(),
nn.MaxPool2d(kernel_size = 2,stride = 2),
nn.Conv2d(128,256,kernel_size = 3,padding = 1),
nn.ReLU(),
nn.Conv2d(256,256,kernel_size=3,padding=1),
nn.ReLU(),
nn.Conv2d(256,256,kernel_size=3,padding=1),
nn.ReLU(),
nn.MaxPool2d(kernel_size = 2,stride = 2),
nn.Conv2d(256,512,kernel_size = 3,padding = 1),
nn.ReLU(),
nn.Conv2d(512,512,kernel_size=3,padding=1),
nn.ReLU(),
nn.Conv2d(512,512,kernel_size=3,padding=1),
nn.ReLU(),
) self.feature_2 = nn.Sequential(
nn.MaxPool2d(kernel_size = 2,stride = 2),
nn.Conv2d(512,512,kernel_size = 3,padding = 1),
nn.ReLU(),
nn.Conv2d(512,512,kernel_size=3,padding=1),
nn.ReLU(),
nn.Conv2d(512,512,kernel_size=3,padding=1),
nn.ReLU(),
)
self.pre = nn.Conv2d(512,64,kernel_size = 1)
self.fc = nn.Conv2d(512,64,kernel_size = 1)
self.det1 = RFBsModule(out = 64,stride = 1)
self.det2 = RFBModule(out = 64,stride = 1)
self.det3 = RFBModule(out = 64,stride = 2)
self.det4 = RFBModule(out = 64,stride = 2)
self.det5 = nn.Conv2d(64,64,kernel_size = 3)
self.det6 = nn.Conv2d(64,64,kernel_size=3) def forward(self,x):
x = self.feature_1(x)
det1 = self.det1(self.fc(x))
x = self.feature_2(x)
x = self.pre(x)
det2 = self.det2(x)
det3 = self.det3(det2)
det4 = self.det4(det3)
det5 = self.det5(det4)
det6 = self.det6(det5)
det1 = det1.permute(0,2,3,1).contiguous().view(x.size(0),-1,64)
det2 = det2.permute(0,2,3,1).contiguous().view(x.size(0),-1,64)
det3 = det3.permute(0,2,3,1).contiguous().view(x.size(0),-1,64)
det4 = det4.permute(0,2,3,1).contiguous().view(x.size(0),-1,64)
det5 = det5.permute(0,2,3,1).contiguous().view(x.size(0),-1,64)
det6 = det6.permute(0,2,3,1).contiguous().view(x.size(0),-1,64) return torch.cat([det1,det2,det3,det4,det5,det6],dim = 1) if __name__ == "__main__":
net = RFBNet()
x = torch.randn(2,3,300,300)
summary(net,(3,300,300),device = "cpu")
print(net(x).size())

论文原文:https://arxiv.org/pdf/1711.07767.pdf

05-15 06:13