YOLOv3训练自己的数据

文章于2020年已更新

https://www.jianshu.com/p/2206db894b28

工具准备

Darknet-YOLO:https://pjreddie.com/darknet/yolo/
labelImg:https://github.com/tzutalin/labelImg

创建文件夹

在darknet/scripts目录下创建以下目录

├── VOCdevkit
│   └── VOC2007
│       ├── Annotations
│       │   ├── 0a0a0b1a-7c39d841.xml
│       │   └── lena.xml
│       ├── ImageSets
│       │   ├── Layout
│       │   ├── Main
│       │   │   ├── test.txt
│       │   │   ├── train.txt
│       │   │   └── val.txt
│       │   └── Segmentation
│       ├── JPEGImages
│       │   ├── 0a0a0b1a-7c39d841.jpg
│       │   └── lena.jpg
│       └── labels
│           └── 0a0a0b1a-7c39d841.txt
└── voc_label.py

其中
JPEGImages下为训练测试集图片

Annotations下为VOC格式的xml标注
如

<annotation>
    <folder>JPEGImages</folder>
    <filename>0a0a0b1a-7c39d841.jpg</filename>
    <path>/home/dew/CV2018/yolo/darknet/scripts/VOCdevkit/VOC2007/JPEGImages/0a0a0b1a-7c39d841.jpg</path>
    <source>
        <database>Unknown</database>
    </source>
    <size>
        <width>1280</width>
        <height>720</height>
        <depth>3</depth>
    </size>
    <segmented>0</segmented>
    <object>
        <name>car</name>
        <pose>Unspecified</pose>
        <truncated>0</truncated>
        <difficult>0</difficult>
        <bndbox>
            <xmin>557</xmin>
            <ymin>275</ymin>
            <xmax>688</xmax>
            <ymax>398</ymax>
        </bndbox>
    </object>
    <object>
        <name>car</name>
        <pose>Unspecified</pose>
        <truncated>0</truncated>
        <difficult>0</difficult>
        <bndbox>
            <xmin>160</xmin>
            <ymin>297</ymin>
            <xmax>252</xmax>
            <ymax>373</ymax>
        </bndbox>
    </object>
    <object>
        <name>car</name>
        <pose>Unspecified</pose>
        <truncated>0</truncated>
        <difficult>0</difficult>
        <bndbox>
            <xmin>392</xmin>
            <ymin>298</ymin>
            <xmax>459</xmax>
            <ymax>353</ymax>
        </bndbox>
    </object>
    <object>
        <name>car</name>
        <pose>Unspecified</pose>
        <truncated>0</truncated>
        <difficult>0</difficult>
        <bndbox>
            <xmin>492</xmin>
            <ymin>304</ymin>
            <xmax>523</xmax>
            <ymax>345</ymax>
        </bndbox>
    </object>
</annotation>

Main下txt文件为对应的测试、训练文件名称
如：

0a0a0b1a-7c39d841

转换标注集格式

修改voc_label.py, 如只有一个class:car

sets=[('2007', 'train'), ('2007', 'val'), ('2007', 'test')]

classes = ["car"]
'''
classes = ["aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike", "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"]
'''

运行文件script/voc_label.py

python ./voc_label.py

会在目录下生成一系列文件并将VOC格式标注转为YOLO格式txt标注（归一化处理）见/darknet/scripts/VOCdevkit/VOC2007/labels/0a0a0b1a-7c39d841.txt

0 0.485546875 0.465972222222 0.10234375 0.170833333333
0 0.16015625 0.463888888889 0.071875 0.105555555556
0 0.331640625 0.450694444444 0.05234375 0.0763888888889
0 0.395703125 0.449305555556 0.02421875 0.0569444444444

修改cfg/voc.data

classes= 1
train  = /home/dew/Desktop/CV2018/yolo/darknet/scripts/2007_train.txt
valid  = /home/dew/Desktop/CV2018/yolo/darknet/scripts/2007_val.txt
names = data/voc.names
backup = backup

修改cfg/yolov3-voc.cfg

查找带有[convolutional]以及[yolo]标签处（共3处）
修改

classes  = 标注种类数
filters=3*(classes+1+4)
ramdom=0  //显存足够1，不足够0

修改data/voc.names

备份后将内容修改为训练集classes名

下载预训练权重文件(只包含卷积层)并训练

wget https://pjreddie.com/media/files/darknet53.conv.74

./darknet detector train cfg/voc.data cfg/yolov3-voc.cfg darknet53.conv.74

log说明

Region xx: cfg文件中yolo-layer的索引；

Avg IOU:当前迭代中，预测的box与标注的box的平均交并比，越大越好，期望数值为1；

Class: 标注物体的分类准确率，越大越好，期望数值为1；

obj: 越大越好，期望数值为1；

No obj: 越小越好；

.5R: 以IOU=0.5为阈值时候的recall; recall = 检出的正样本/实际的正样本

0.75R: 以IOU=0.75为阈值时候的recall;

count:正样本数目。