readme.md · monkeyfx/detection

目标识别-图像处理中的霸王龙
===
# 1. 项目总览
在这个项目中，我们实现了SSD，RFB，FSSD，YOLOV3，YOLOV4以及YoloV3+ASFF。硬件使用的是Tesla P4 GPU

# 2.物体检测总体指标(SSD, RFB与FSSD)
V = PASCAL VOC 2012+2014

C = MS COCO 2017

## 2.1.SSD
### 2.1.1.VGG16
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | [76.2%](mAP/ssd/voc/vgg16.txt) :+1: | - | - | [79.2%](mAP/ssd512/voc/vgg16.txt):+1: | - | - |
| V | ![Images](results/ssd/VOC/vgg16_sn_300.jpg) | - | - | ![Images](results/ssd/VOC/vgg16_sn_512.jpg) | - | - |
| C | - | - | - | - | - | - |
| C | - | - | - | - | - | - |

### 2.1.2.Resnet50
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | [80.1%](mAP/ssd/voc/resnet50.txt):+1: | - | - | - | - | - |
| V | ![Images](results/ssd/VOC/resnet50_sn_300.jpg) | - | - | - | - | - |
| C | [25.0%](mAP/ssd/coco/resnet50.txt) :+1: | - | - | - | - | - |
| C | ![Images](results/ssd/COCO/resnet50_sn_300.jpg) | - | - | - | - | - |

### 2.1.3.Resnet152
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | [73.4%](mAP/ssd/voc/resnet152.txt) :+1: | - | - | - | - | - |
| V | ![Images](results/ssd/VOC/resnet152_sn_300.jpg) | - | - | - | - | - |
| C | - | - | - | - | - | - |
| C | - | - | - | - | - | - |

### 2.1.4.Darknet19
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | [74.3%](mAP/ssd/voc/darknet19.txt):+1: | - | - | - | - | - |
| V | ![Images](results/ssd/VOC/darknet19_sn_300.jpg) | - | - | - | - | - |
| C | [20.5%](mAP/ssd/coco/darknet19.txt) :+1: | - | - | - | - | - |
| C | ![Images](results/ssd/COCO/darknet19_sn_300.jpg) | - | - | - | - | - |

### 2.1.5.mobilenetv1
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | [72.7%](mAP/ssd/voc/mobilenetv1.txt):+1: | - | - | - | - | - |
| V | ![Images](results/ssd/VOC/mobilenetv1_sn_300.jpg) | - | - | - | - | - |
| C | [18.8%](mAP/ssd/coco/mobilenetv1.txt) :+1: | - | - | - | - | - |
| C | ![Images](results/ssd/COCO/mobilenetv1_sn_300.jpg) | - | - | - | - | - |

## 2.2.RFB
### 2.2.1.VGG16
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | [79.0%](mAP/rfb/voc/vgg16.txt):+1: | - | - | [79.7%](mAP/rfb512/voc/vgg16.txt):+1: | - | - |
| V | ![Images](results/rfb/VOC/vgg16_sn_300.jpg) | - | - | ![Images](results/rfb/VOC/vgg16_sn_512.jpg) | - | - |
| C | - | - | - | - | - | - |
| C | - | - | - | - | - | - |

### 2.2.2.Resnet50
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | - | - | - | - | - | - |
| V | - | - | - | - | - | - |
| C | - | - | - | - | - | - |
| C | - | - | - | - | - | - |

### 2.2.3.Resnet152
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | - | - | - | - | - | - |
| V | - | - | - | - | - | - |
| C | - | - | - | - | - | - |
| C | - | - | - | - | - | - |

### 2.2.4.Darknet19
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | [76.2%](mAP/rfb/voc/darknet19.txt):+1: | - | - | - | - | - |
| V | ![Images](results/rfb/VOC/darknet19_sn_300.jpg) | - | - | - | - | - |
| C | [22.4%](mAP/rfb/coco/darknet19.txt) :+1: | - | - | - | - | - |
| C | ![Images](results/rfb/COCO/darknet19_sn_300.jpg) | - | - | - | - | - |

### 2.2.5.mobilenetv1
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | [73.8%](mAP/rfb/voc/mobilenetv1.txt):+1: | - | - | - | - | - |
| V | ![Images](results/rfb/VOC/mobilenetv1_sn_300.jpg) | - | - | - | - | - |
| C | [19.0%](mAP/rfb/coco/mobilenetv1.txt) :+1: | - | - | - | - | - |
| C | ![Images](results/rfb/COCO/mobilenetv1_sn_300.jpg) | - | - | - | - | - |

## 2.3.FSSD
### 2.3.1.VGG16
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | [77.9%](mAP/fssd/voc/vgg16.txt):+1: | - | - | - | - | - |
| V | ![Images](results/fssd/VOC/vgg16_sn_300.jpg) | - | - | - | - | - |
| C | - | - | - | - | - | - |
| C | - | - | - | - | - | - |

### 2.3.2.Resnet50
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | [74.0%](mAP/fssd/voc/resnet50.txt) :+1: | - | - | - | - | - |
| V | ![Images](results/fssd/VOC/resnet50_sn_300.jpg) | - | - | - | - | - |
| C | [26.6%](mAP/fssd/coco/resnet50.txt) :+1: | - | - | - | - | - |
| C | ![Images](results/fssd/COCO/resnet50_sn_300.jpg) | - | - | - | - | - |

### 2.3.3.Resnet152
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | [74.1%](mAP/fssd/voc/resnet152.txt) :+1: | - | - | - | - | - |
| V | ![Images](results/fssd/VOC/resnet152_sn_300.jpg) | - | - | - | - | - |
| C | - | - | - | - | - | - |
| C | - | - | - | - | - | - |

### 2.3.4.Darknet19
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | [78.2%](mAP/fssd/voc/darknet19.txt):+1: | - | - | - | - | - |
| V | ![Images](results/fssd/VOC/darknet19_sn_300.jpg) | - | - | - | - | - |
| C | [25.2%](mAP/fssd/coco/darknet19.txt) :+1: | - | - | - | - | - |
| C | ![Images](results/fssd/COCO/darknet19_sn_300.jpg) | - | - | - | - | - |

### 2.3.5.mobilenetv1
| - | $300 \times 300$ | $300 \times 300$ | $300 \times 300$ | $512 \times 512$ | $512 \times 512$ | $512 \times 512$ |
| ------- | ------------ | -------- | ------------ | ------------ | -------- | ------------ |
| - | SmoothL1+nms | Diou+nms | Ciou+diounms | SmoothL1+nms | Diou+nms | Ciou+diounms |
| V | [73.5%](mAP/fssd/voc/mobilenetv1.txt):+1: | - | - | - | - | - |
| V | ![Images](results/fssd/VOC/mobilenetv1_sn_300.jpg) | - | - | - | - | - |
| C | [22.8%](mAP/fssd/coco/mobilenetv1.txt) :+1: | - | - | - | - | - |
| C | ![Images](results/fssd/COCO/mobilenetv1_sn_300.jpg) | - | - | - | - | - |

# 3.物体检测总体指标(YOLOV3, YOLOV4, YOLOV3+ASFF)

# 4.总体介绍
每个网络都分为网络结构，数据整理与损失函数三个部分，其它的基本一样。
## 4.1.SSD，RFB与FSSD
### 4.1.1.数据整理
这三个网络在训练数据处理方面是同一个方法，为[SSD.py](ELib/dataset/SSD.py)

### 4.1.2.网络结构
这三大网络，背后的骨干网络Backbone都是一样的，包括VGG16, Resnet50, Resnet152, MobileNetV1, MobileNetV2等等。
不同的是这三个网络对于骨干网络的Feature Maps的获取层数与处理不同。

### 4.1.3.损失函数
这三大网络的损失函数是一样的，为[Multibox.py](ELib/loss/multibox.py)

## 4.2.YOLOV3与YOLOV4

monkeyfx/detection

简介

发行版

贡献者

近期动态

monkeyfx/detection .gitee-modal { width: 500px !important; }

简介

发行版

贡献者

近期动态

搜索帮助

monkeyfx/detection