yolov3

Posted 2021-06-11 crazybird123
tags:
篇首语：本文由小常识网(cha138.com)小编为大家整理，主要介绍了yolov3相关的知识，希望对你有一定的参考价值。
1、darknet53
Backbone
输入: 416x416x3
3x3Conv2d, stride=1, padding=1 -> BatchNorm2d -> LeakyReLu : [1, 32, 416, 416]  

[1,2,8,8,4]五种尺度，一共32倍下采样
layer1:重复1次
3x3Conv2d, stride=2, padding=1 -> BatchNorm2d -> LeakyReLu : [1, 64, 208, 208]   下采样，channel加倍

残差模块
1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 32, 208, 208]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 64, 208, 208]  3x3卷积升维

layer2:重复2次
3x3Conv2d, stride=2, padding=1 -> BatchNorm2d -> LeakyReLu : [1, 128, 104, 104]   下采样，channel加倍
残差模块
1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 64, 104, 104]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 128, 104, 104]  3x3卷积升维
残差模块
1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 64, 104, 104]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 128, 104, 104]  3x3卷积升维

layer3:重复8次 这个特征需要使用，特征图尺寸为52x52x256
3x3Conv2d, stride=2, padding=1 -> BatchNorm2d -> LeakyReLu : [1, 256, 52, 52]   下采样，channel加倍
残差模块
1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 128, 52, 52]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 256, 52, 52]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 128, 52, 52]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 256, 52, 52]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 128, 52, 52]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 256, 52, 52]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 128, 52, 52]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 256, 52, 52]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 128, 52, 52]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 256, 52, 52]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 128, 52, 52]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 256, 52, 52]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 128, 52, 52]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 256, 52, 52]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 128, 52, 52]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 256, 52, 52]  3x3卷积升维


layer4:重复8次 这个特征需要使用，特征图尺寸为26x26x512
3x3Conv2d, stride=2, padding=1 -> BatchNorm2d -> LeakyReLu : [1, 512, 26, 26]   下采样，channel加倍

残差模块
1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 256, 26, 26]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 512, 26, 26]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 256, 26, 26]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 512, 26, 26]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 256, 26, 26]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 512, 26, 26]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 256, 26, 26]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 512, 26, 26]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 256, 26, 26]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 512, 26, 26]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 256, 26, 26]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 512, 26, 26]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 256, 26, 26]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 512, 26, 26]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 256, 26, 26]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 512, 26, 26]  3x3卷积升维

layer5: 重复4次 这个特征需要使用，特征图尺寸为13x13x1024
3x3Conv2d, stride=2, padding=1 -> BatchNorm2d -> LeakyReLu : [1, 1024, 13, 13]   下采样，channel加倍

残差模块
1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 512, 13, 13]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 1024, 13, 13]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 512, 13, 13]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 1024, 13, 13]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 512, 13, 13]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 1024, 13, 13]  3x3卷积升维

1x1Conv2d, stride=1, padding=0 -> BatchNorm2d -> LeakyReLu : [1, 512, 13, 13]   1x1卷积降维
3x3Conv2d, stride=1, padding=1  -> BatchNorm2d -> LeakyReLu : [1, 1024, 13, 13]  3x3卷积升维

Head
yolo branch 0  ( out0 )
out0_branch----x0: (1x1024x13x13) -> 1x512x13x13 -> 1x1024x13x13 -> 1x512x13x13 -> 1x1024x13x13 -> 1x512x13x13
out0------------------out0_branch -> 1x1024x13x13 -> 1x计算出来的数x13x13

 yolo branch 1
 1x512x13x13 -> 1x256x13x13  使用1x1卷积调整通道数
out0_branch (1x256x13x13)-> 1x256x26x26 上采样
 1x256x26x26 + 1x512x26x26 -> 1x768x26x26  特征图融合
 out1_branch ---------- x1_in: 1x768x26x26 -> 1x256x26x26 -> 1x512x26x26 -> 1x256x26x26 -> 1x512x26x26 ->1x256x26x26
 out1 --------------------------out1_branch -> 1x512x26x26 ->1x计算出来的数x26x26

yolo branch 2  
26,26,256 -> 26,26,128
26,26,128 -> 52,52,128 上采样
1x128x52x52  + 1x256x52x52  -> 1x384x52x52  融合
#out2----1x384x52x52 -> 1x128x52x52 -> 1x256x52x52 -> 1x128x52x52 -> 1x256x52x52 -> 1x128x52x52 -> 1x256x52x52 -> 1X计算出来的数X52X52



损失函数
GroundTruth： images-8x3x416x416,  8-batchsize  3-channel  416-height 416-width
                        targets-8x(n1+n2+....+n8)x5---8张图，每张图有nk个真值框
网络输出：outputs---3x(8x(5+类别数)x13x13)    3x(8x(5+类别数)x26x26)     3x(8x(5+类别数)x52x52)

第一张特征图：3x(8x(5+类别数)x13x13)  检测大物体
预测框中心点坐标x：【8,3,13,13】 sigmoid
预测框中心点坐标y：【8,3,13,13】sigmoid
预测框宽度w：         【8,3,13,13】
预测框高度h:            【8,3,13,13】
预测框有没有物体conf:【8,3,13,13】sigmoid
预测框类别预测：    【8,3,13,13，6】 6是类别数 有6类物体需要识别 sigmoid

重点1---找到哪些先验框内部包含物体
（1）计算GroundTruth在13x13特征图上的坐标x,y,w,h，向下取整看x,y落在哪个方格
  (2)预测框和anchor左上角移动至坐标原点，利用宽高算IOU
  (3)计算框中心点坐标x，y距离所属的单元格左上角偏移量，宽度w和高度h相对于特征图13x13中的anchor宽度和高度进行编码

重点2---将预测结果进行解码，判断预测结果和真实值的重合程度
以上是关于yolov3的主要内容，如果未能解决你的问题，请参考以下文章
matlab能否运行yolov3
从零开始学习YOLOv37. 教你在YOLOv3模型中添加Attention机制
TensorFlow Lite - 对象检测 API YOLOv3
YOLOV3简介
收藏 | Yolov3和Yolov4核心内容代码梳理