Пользовательский набор данных Yolo для обучения

компьютерное зрение

Создание среды Yolov4

Окружение здесь примерно такое же, как и у yolov3, разница в основном в предпоездных весах и условных соединениях.

Cloning and Building Darknet

clone darknet from AlexeyAB's famous repository,

git clone https://github.com/AlexeyAB/darknet

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

build darknet

make

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

Test env Enabled

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

image.png

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

  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

    [{
        "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
            }
        ]
    },
    //...
]

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)

Пользовательский набор данных для обучения Yolov4

Общий метод тот же, что и у yolov3, но на вчерашнем обучении использовались более зрелые решения, и конкретные шаги не до конца понятны, поэтому сегодня я решил еще раз подробно изучить с нуля.

  • 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

Потому что задачи лаборатории специфичны для нескольких видов игрушек, и жестких требований по масштабируемости нет. Таким образом, набор данных с открытым исходным кодом Google используется только для обучения, или метод маркировки набора данных сам по себе используется для создания набора данных.

Manually Labeling Images with labelImg(Annotation Tool)


На данный момент подготовлены наборы данных для обучения и проверки.

Configuring Files for Training

cfg file

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

  • bash=64 & subdivisions=16: Параметры, рекомендованные в Интернете

    Здесь ограничена мощность сервера, установленная SubDivision 32, но скорость все равно очень низкая

  • 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

obj.data

classes= 4
train  = data/train.txt
valid  = data/test.txt
names = data/obj.names
backup = backup
  • 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

train

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

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

the highest mAP, the most accurate is

Категории