论文:Receptive Field Block Net for Accurate and Fast Object Detection
Github:https://github.com/ruinmessi/RFBNet
ECCV2018
论文基于Receptive Fields (RFs) ,提出了RF Block (RFB) ,然后基于RFB,对以VGG16为基础架构的ssd进行了改进,形成最终的RFBNet。RFBNet是一种one-stage的架构,精度高于传统ssd,速度较传统ssd略慢。
论文贡献:
- 提出RFB模块
- 提出了RFBNet
- RFBNet在voc和coco上都取得了实时的处理速度。并且将VGG网络换为MobileNet也具有性能提升,证明RFB模块的泛化性。
RFB流程:
RFB模块分为3个分支,每个分支的底层都经过不同大小的卷积核进行处理(1_1,3_3,5_5),然后经过3_3空洞卷积,膨胀率分别为(1,3,5)。这样使得不同分支都具有不同的感受野,最后将所有的feature map连接起来。
Inception,ASPP,Deformable conv,RFB对比:
可以看出Inception,ASPP,Deformable conv,这3种方法的底层输入都是同样感受野的feature map。而RFB的底层输入具有不同的感受野。
RFB算是Inception和ASPP的综合,具有更大的感受野。
RFB模块:
RFB模块分为RFB,RFB-s两种。程序和实现和上面的图示,有略微的区别。
-
class BasicRFB(nn.Module):
-
-
def init(self, in_planes, out_planes, stride=1, scale = 0.1, visual = 1):
-
super(BasicRFB, self).__init__()
-
self.scale = scale
-
self.out_channels = out_planes
-
inter_planes = in_planes // 8
-
self.branch0 = nn.Sequential(
-
BasicConv(in_planes, 2_inter_planes, kernel_size=1, stride=stride),
-
BasicConv(2_inter_planes, 2_inter_planes, kernel_size=3, stride=1, padding=visual, dilation=visual, relu=False)
-
)
-
self.branch1 = nn.Sequential(
-
BasicConv(in_planes, inter_planes, kernel_size=1, stride=1),
-
BasicConv(inter_planes, 2_inter_planes, kernel_size=(3,3), stride=stride, padding=(1,1)),
-
BasicConv(2_inter_planes, 2_inter_planes, kernel_size=3, stride=1, padding=visual+1, dilation=visual+1, relu=False)
-
)
-
self.branch2 = nn.Sequential(
-
BasicConv(in_planes, inter_planes, kernel_size=1, stride=1),
-
BasicConv(inter_planes, (inter_planes//2)_3, kernel_size=3, stride=1, padding=1),
-
BasicConv((inter_planes//2)_3, 2_inter_planes, kernel_size=3, stride=stride, padding=1),
-
BasicConv(2_inter_planes, 2_inter_planes, kernel_size=3, stride=1, padding=2_visual+1, dilation=2_visual+1, relu=False)
-
)
-
-
self.ConvLinear = BasicConv(6_inter_planes, out_planes, kernel_size=1, stride=1, relu=False)
-
self.shortcut = BasicConv(in_planes, out_planes, kernel_size=1, stride=stride, relu=False)
-
self.relu = nn.ReLU(inplace=False)
-
-
def forward(self,x):
-
x0 = self.branch0(x)
-
x1 = self.branch1(x)
-
x2 = self.branch2(x)
-
-
out = torch.cat((x0,x1,x2),1)
-
out = self.ConvLinear(out)
-
short = self.shortcut(x)
-
out = out*self.scale + short
-
out = self.relu(out)
-
-
return out
RFB-Net :
实验结果:
RFB和其他基础骨架网络在VOC的对比结果
RFB各模块有效性
RFB和其他增加感受野模块对比
RFB和其他基础骨架网络在coco的对比结果
RFB速度对比
总结:
RFB模块是一种增加感受野的模块,基于此改进的RFB-Net 具有速度和精度的双重优势。
评论(0)
您还未登录,请登录后发表或查看评论