Yolo训练自定义数据集

Yolov4环境搭建

这里的环境与yolov3大致相同，差别主要在pre-train weights和conv连接

Cloning and Building Darknet

clone darknet from AlexeyAB’s famous repository,

git clone https://github.com/AlexeyAB/darknet
复制代码

adjust the Makefile to enable OPENCV and GPU for darknet

cd darknet
sed -i 's/OPENCV=0/OPENCV=1/' Makefile
sed -i 's/GPU=0/GPU=1/' Makefile
sed -i 's/CUDNN=0/CUDNN=1/' Makefile
sed -i 's/CUDNN_HALF=0/CUDNN_HALF=1/' Makefile
复制代码

build darknet

make
复制代码

Pre-trained yolov4 weights

YOLOv4 has been trained already on the coco dataset which has 80 classes that it can predict.

wget https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4.weights
复制代码

Test env Enabled

./darknet detector test cfg/coco.data cfg/yolov4.cfg yolov4.weights data/person.jpg
复制代码

Multiple Images at Once

1. make a .txt file which has the paths to several images want to be detected at once

   data/person.jpg
   data/horses.jpg
   data/giraffe.jpg
   data/dog.jpg
   复制代码

2. save result to .json file

   ./darknet detector test cfg/coco.data cfg/yolov3.cfg yolov3.weights -ext_output -dont_show -out result.json < list.txt
   复制代码

   [{
           "frame_id": 1,
           "filename": "data/person.jpg",
           "objects": [{
                   "class_id": 17,
                   "name": "horse",
                   "relative_coordinates": {
                       "center_x": 0.783550,
                       "center_y": 0.566949,
                       "width": 0.335207,
                       "height": 0.486880
                   },
                   "confidence": 0.997604
               },
               {
                   "class_id": 16,
                   "name": "dog",
                   "relative_coordinates": {
                       "center_x": 0.206590,
                       "center_y": 0.722102,
                       "width": 0.229715,
                       "height": 0.210134
                   },
                   "confidence": 0.994348
               },
               {
                   "class_id": 0,
                   "name": "person",
                   "relative_coordinates": {
                       "center_x": 0.364771,
                       "center_y": 0.558493,
                       "width": 0.134738,
                       "height": 0.669826
                   },
                   "confidence": 0.999949
               }
           ]
       },
       //...
   ]
   复制代码

Yolo command line flags

• -thresh: add a threshold for confidences on the detections, only detections with a confidence level above the threshold will be returned

• -dont_show: not have the image outputted after running darknet

• -ext_output: output bounding box coordinates

  dog: 99%	(left_x:   59   top_y:  262   width:  147   height:   89)
  person: 100%	(left_x:  190   top_y:   95   width:   86   height:  284)
  horse: 100%	(left_x:  394   top_y:  137   width:  215   height:  206)
  复制代码

Yolov4训练自定义数据集

大致方法与yolov3相同，但昨天的训练更多的使用比较成熟的解决方案，对具体步骤还不是完全明白掌握，因此决定今天再次详细的从头进行学习

• Labeled Custom Dataset
• Custom .cfg file
• obj.data and obj.names files
• train.txt file (test.txt is optional here as well)

Gathering and Labeling a Custom Dataset

Using Google’s Open Images Dataset

由于实验室的任务针对特定几种玩具，且在扩展性上没有太强硬的要求。因此谷歌的开源数据集仅做学习使用，还是采用自己标注数据集的方法进行数据集的构建

• Google’s Open Images Dataset
• OIDv4 toolkit

Manually Labeling Images with labelImg(Annotation Tool)

• labelImg工具使用

至此已经准备好用于train和valid的数据集了

Configuring Files for Training

cfg file

edit the yolov4.cfg to fit the needs based on the object detector

• bash=64 & subdivisions=16：网上比较推荐的参数

  这里受限于服务器容量 将subdivisions设为32，但是速度仍然很慢

• classes=4 in the three YOLO layers

• filters=(classes + 5) * 3: three convolutional layers before the YOLO layers

• width=416 & height=416: any multiple of 32, 416 is standard

  • improve results by making value larger like 608 but will slow down training

• max_batches=(# of classes) * 2000: but no less than 6000

• steps=(80% of max_batches), (90% of max_batches)

• random=1: if run into memory issues or find the training taking a super long time, change three yolo layers from 1 to 0 to speed up training but slightly reduce accurancy of model

obj.names

one class name per line in the same order as dataset generation step

NOTE: don’t have spaces in class name, use _ for replacement

sheep
giraffe
cloud
snow
复制代码

obj.data

classes= 4
train  = data/train.txt
valid  = data/test.txt
names = data/obj.names
backup = backup
复制代码

• backup: where save the weights to of the model throughout training

train.txt and test.txt

hold the reletive paths to all the training images and valididation images, it contain one line for each training image path or validation image path

Train Custom Object Detector

Download pre-trained weights for the convolutional layers. By using these weights it helps custom object detector to be way more accurate and not have to train as long.

wget https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4.conv.137
复制代码

train

./darknet detector train ../../data/obj.data cfg/yolov4_custom.cfg yolov4.conv.137 -dont_show
复制代码

Checking the mAP of the Model

mAP: mean average precision

./darknet detector map ../../data/obj.data cfg/yolov4_custom.cfg backup/yolov4_custom_last.weights
复制代码

the highest mAP, the most accurate is