SSD算法学习及PyTorch代码分析[1]-整体框架

SSD(Single Shot Multibox Detector)是one-stage目标检测算法的典型代表，SSD在速度上表现不错，精度上也不差，是一个非常优秀的算法。

这里，通过SSDPyTorch代码进行分析学习。这篇主要分析SSD的整体网络，有个大致的概念。
一些用到的卷积计算公式：

图像卷积输出大小公式(正常)： $o=\lfloor \frac{i-k+2p}{s}\rfloor +1.$o=⌊si−k+2p​⌋+1.

图像卷积输出大小公式(ceil_mode)： $o=\lceil \frac{i-k+2p}{s}\rceil +1.$o=⌈si−k+2p​⌉+1.

图像卷积输出大小公式(带空洞卷积 $d$d)： $o=\lceil \frac{i-k+2p-(k-1)\ast (d-1)}{s}\rceil +1.$o=⌈si−k+2p−(k−1)∗(d−1)​⌉+1.

$i$i为输入图片大小， $k$k为卷积核大小， $p$p为padding大小， $s$s为stride大小, $d$d为(空格数+1)

1. VGG部分 {conv1_2, conv2_2, conv3_3, conv4_3, conv5_3, fc6(conv6), fc7(conv7)}

# 这里给出输入图像的大小（C,H,W）
input_size:(3, 300, 300)
# conv1_2
Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace)
Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace)
MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
# 这里给出通过conv_2后图像计算方式和大小， 后面的image_size亦是如此
image_size:(300-2+2*0)/2+1=150 (64, 150, 150)
    
# conv2_2
Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace)
Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace)
MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
image_size:(150-2+2*0)/2+1=75 (128, 75, 75)
    
# conv3_3
Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace)
Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace)
Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace)
MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=True)
image_size: ceil[(75-2+2*0)/2+1]=38 (256, 38, 38)
   
# conv4_3
Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace)
Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace)
Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))#-->
ReLU(inplace)
MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
image_size:(38-2+2*0)/2+1=19 (512, 19, 19)
    
# conv5_3
Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace)
Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace)
Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace)
MaxPool2d(kernel_size=3, stride=1, padding=1, dilation=1, ceil_mode=False)
image_size:(19-3+2*1)/1+1=75 (64, 19, 19)

# conv6,空洞卷积
Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(6, 6), dilation=(6, 6))
ReLU(inplace)
image_size:(19-3+2*6-(3-1)*(6-1)/1+1=19 (1024, 19, 19)
    
# conv7
Conv2d(1024, 1024, kernel_size=(1, 1), stride=(1, 1))#-->
ReLU(inplace)
image_size:(19-1+2*0)/1+1=19 (1024, 19, 19)

2. Extra Feature Layers{conv8_2, conv9_2, conv10_2, conv11_2}

input_size:(19,19)
# conv8_2
Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1))
Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1)) #-->
image_size:(19-3+2*1)/2+1=10 (10,10)

# conv9_2
Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1))
Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))#-->
image_size: (10-3+2*1)/2+1=5 (5,5)
    
# conv10_2
Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1))
Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1))#-->
image_size: (5-3+2*0)/1+1=3 (3,3)

# conv11_2
Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1))
Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1))#-->
image_size: (3-3+2*0)/2+1=1 (1,1)

其中#-->表示连接到detections层，做定位与置信度分类层

3. Loc Layer

Conv2d(512, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Conv2d(1024, 24, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Conv2d(512, 24, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Conv2d(256, 24, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Conv2d(256, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))

4. Conf Layer

Conv2d(512, 84, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Conv2d(1024, 126, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Conv2d(512, 126, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Conv2d(256, 126, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Conv2d(256, 84, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
Conv2d(256, 84, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))