https://github.com/dataxujing/diou_yolo_v3
:chart_with_upwards_trend::chart_with_upwards_trend::chart_with_upwards_trend:【口罩佩戴检测数据训练 | 开源口罩检测数据集和预训练模型】Train D/CIoU_YOLO_V3 by darknet for object detection
https://github.com/dataxujing/diou_yolo_v3
c ciou cv darknet diou object-detection yolov3
Last synced: about 1 month ago
JSON representation
:chart_with_upwards_trend::chart_with_upwards_trend::chart_with_upwards_trend:【口罩佩戴检测数据训练 | 开源口罩检测数据集和预训练模型】Train D/CIoU_YOLO_V3 by darknet for object detection
- Host: GitHub
- URL: https://github.com/dataxujing/diou_yolo_v3
- Owner: DataXujing
- License: gpl-3.0
- Created: 2019-12-19T14:42:19.000Z (almost 6 years ago)
- Default Branch: FaceMask
- Last Pushed: 2020-04-03T02:55:09.000Z (over 5 years ago)
- Last Synced: 2025-08-16T15:41:12.770Z (2 months ago)
- Topics: c, ciou, cv, darknet, diou, object-detection, yolov3
- Language: Jupyter Notebook
- Homepage:
- Size: 27.5 MB
- Stars: 58
- Watchers: 2
- Forks: 17
- Open Issues: 4
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
## [DIoU-Darknet](https://github.com/Zzh-tju/DIoU-darknet) 口罩佩戴检测
## 提供整理的开源口罩佩戴检测数据集的下载！！！
+ 预训练的模型： 可下载在网上找图像或视频测试！
+ 训练数据下载地址：百度云盘链接： 提取码：h9sl
## 检测效果 **好于**[AIZOO](https://aizoo.com/face-mask-detection.html)
## 检测效果 **好于**[百度开源口罩检测](https://www.paddlepaddle.org.cn/hub/scene/maskdetect)
> YOLOv3 with DIoU and CIoU losses implemented in Darknet
[[arxiv](https://arxiv.org/abs/1911.08287)] [[pdf](https://arxiv.org/pdf/1911.08287.pdf)]
**Xu Jing**
**DEMO**
相同图片基于AIZOO开源的检测效果:
配置和训练过程以darknet版YOLO v3和Gaussian YOLO v3的配置训练过程相似，这里以Ubuntu为例，关于Windows下的配置和生成过程，具体可以联系[**DataXujing**](https://github.com/DataXujing)
### 1.修改Makefile并make
```shell
GPU=1 #如果使用GPU设置为1，CPU设置为0
CUDNN=1 #如果使用CUDNN设置为1，否则为0
OPENCV=1 #如果调用摄像头，还需要设置OPENCV为1，否则为0
OPENMP=1 #如果使用OPENMP设置为1，否则为0
DEBUG=0 #如果使用DEBUG设置为1，否则为0
```
```shell
make
```
### 2.数据来源
+ 1. [AIZOO](https://aizoo.com/face-mask-detection.html)
+ 2. 知乎：阿扣是谁哇
+ 3. 个人爬取的部分数据及标注，标注工具使用了[LabelImg](https://github.com/tzutalin/labelImg)
数据整理成VOC格式的数据，最终数据为8565张图像，由于存储空间问题，如果开发者需要改数据，请通过个人GitHub主页公布的邮箱索要。
将数据存放在`data/myData`下
### 3.配置文件
+ (1).Losses
目前支持的损失：`[iou|giou|diou|ciou|mse]`
```
iou_loss=ciou
```
+ (2).Normalizers
Location和Claffication损失之间的Normalize,`cls_normalizer` 和`iou_normalizer` 默认都是`1.0`,但在DIoU探索中发现，下面的配置比较好：
```
iou_loss=diou
cls_normalizer=1
iou_normalizer=1.0
```
```
iou_loss=ciou
cls_normalizer=1
iou_normalizer=0.5
```
+ (3).DIoU-NMS
在`nms_kind`参数下设置，目前支持`[greedynms|diounms]`
```
nms_kind=greedynms
```
```
nms_kind=diounms
```
对于YOLOv3来说, DIoU-NMS中引入beta1参数： `DIoU = IoU - R_DIoU ^ {beta1}`,是的NMS表现不错。DIoU发现，对于YOLOv3
```
beta1=0.6
```
DIoU-NMS是比较好的(bet1=1.0也优于greedy-NMS)。但对于SSD和Faster R-CNN而言，`beta1=1.0`就很不错了。
+ (4).learning rate
- `learning_rate`的设置原则 `NEW_RATE = ORIGINAL_RATE * 1/NUMBER_OF_GPUS` ；
- `burn_in` 的设置原则 `NEW_BURN_IN = ORIGINAL_BURN_IN * NUMBER_OF_GPUS`
因此对于1个GPU的情形：
```
learning_rate=0.001
burn_in=1000
```
对于2个GPU的情形:
```
learning_rate=0.0005
burn_in=2000
```
对于4个GPU的情形:
```
learning_rate=0.00025
burn_in=4000
```
+ (5).其他配置
其他配置和YOLOv3相似！
### 4.训练
```
./darknet detector train cfg/myData-diou.data cfg/myData-diou.cfg darknet53.conv.74 #-gpus 0,1
```
### 5.Evaluation
./darknet detector test cfg/myData-ciou.data cfg/myData-ciou.cfg backup/ciou/myData-ciou_14000.weights data/test1.jpg
### Reference
```
@inproceedings{zheng2020distance,
author = {Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, Dongwei Ren},
title = {Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression},
booktitle = {The AAAI Conference on Artificial Intelligence (AAAI)},
year = {2020},
}
```