yolo-studio

A turnkey solution to train and deploy your own object detection network, contains:

• Augmentor - image augmentation library in Python.
• Yolo_mark - the toolkit to prepare training data.
• darknet - the main engine for training & inferencing.
• yolo2_light - lightweighted inferencing engine, optional.

How to build

Install CUDA

• Uninstall Geforce Experience and current driver

• CUDA 9.1: https://developer.nvidia.com/cuda-downloads

• cuDNN v7.x for CUDA: https://developer.nvidia.com/rdp/cudnn-download

  • Extract to the same folder as CUDA SDK
  • e.g. c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.1\

Install OpenCV

• OpenCV 3.4.0: https://sourceforge.net/projects/opencvlibrary/files/opencv-win/3.4.0/opencv-3.4.0-vc14_vc15.exe/download
• Extract to d:\opencv\
• Make symbolic link from c:\opencv_3.0\opencv\ to d:\opencv\

Build from source

Execute the batch file

build.bat

Or build the componets from Visual Studio

• darknet: darknet\build\darknet\darknet.sln, x64|Release -> darknet\build\darknet\x64\darknet.exe
• Yolo_mark: Yolo_mark\yolo_mark.sln, x64|Release -> Yolo_mark\x64\Release\yolo_mark.exe
• yolo2_light: yolo2_light\yolo_gpu.sln, Release -> yolo2_light\bin\yolo_gpu.exe

Fine tune a existing network

https://github.com/pjreddie/darknet/wiki/YOLO:-Real-Time-Object-Detection

darknet.exe partial cfg/darknet19_448.cfg darknet19_448.weights darknet19_448.conv.23 23

Object Detection - yolo

How to mark labelled images

• delete all files from directory my-yolo-net/img and put your .jpg-images in
• change numer of classes (objects for detection) in file my-yolo-net/obj.data: https://github.com/jing-vision/yolo-studio/blob/master/networks/yolo-template/obj.data#L1
• put names of objects, one for each line in file my-yolo-net/obj.names: https://github.com/jing-vision/yolo-studio/blob/master/networks/yolo-template/obj.names
• Run file: my-yolo-net/yolo_mark.cmd

Train yolo v2

0. Fork networks/yolov2-template to networks/my-yolo-net

1. Download pre-trained weights for the convolutional layers: http://pjreddie.com/media/files/darknet19_448.conv.23 to bin/darknet19_448.conv.23

2. To training for your custom objects, you should change 2 lines in file obj.cfg:

• change classes in obj.data#L1
• set number of classes (objects) in obj.cfg#L230
• set filter-value equal to (classes + 5)*5 in obj.cfg#L224

3. Run my-yolo-net/train.cmd

Train yolo v3

0. Fork networks/yolov3-template to networks/my-yolo-net

1. Download pre-trained weights for the convolutional layers: http://pjreddie.com/media/files/darknet53.conv.74 to bin/darknet53.conv.74

2. Create file obj.cfg with the same content as in yolov3.cfg (or copy yolov3.cfg to obj.cfg) and:

• change line batch to batch=64
• change line subdivisions to subdivisions=8
• change line classes=80 to your number of objects in each of 3 [yolo]-layers:
  • obj.cfg#L610
  • obj.cfg#L696
  • obj.cfg#L783
• change [filters=255] to filters=(classes + 5)x3 in the 3 [convolutional] before each [yolo] layer
  • obj.cfg#L603
  • obj.cfg#L689
  • obj.cfg#L776

So if classes=1 then should be filters=18. If classes=2 then write filters=21.

(Do not write in the cfg-file: filters=(classes + 5)x3)

(Generally filters depends on the classes, coords and number of masks, i.e. filters=(classes + coords + 1)*<number of mask>, where mask is indices of anchors. If mask is absence, then filters=(classes + coords + 1)*num)

So for example, for 2 objects, your file obj.cfg should differ from yolov3.cfg in such lines in each of 3 [yolo]-layers:

[convolutional]
filters=21

[region]
classes=2

How to inference

Pre-trained models for different cfg-files can be downloaded from (smaller -> faster & lower quality):

cfg	weights
cfg/yolov2.cfg	https://pjreddie.com/media/files/yolov2.weights
cfg/yolov2-tiny.cfg	https://pjreddie.com/media/files/yolov2-tiny.weights
cfg/yolo9000.cfg	http://pjreddie.com/media/files/yolo9000.weights
cfg/yolov3.cfg	https://pjreddie.com/media/files/yolov3.weights
cfg/yolov3-tiny.cfg	https://pjreddie.com/media/files/yolov3-tiny.weights

Run Darknet

General

darknet.exe detector demo <data> <cfg> <weights> -c <camera_idx>
darknet.exe detector demo <data> <cfg> <weights> <video_filename>
darknet.exe detector test <data> <cfg> <weights> <img_filename>

Default launch device combination is -i 0 -c 0.

Run from bin/ folder

yolov3

darknet.exe detector demo cfg/coco.data cfg/yolov3.cfg yolov3.weights

yolo9000 on camera #0

darknet.exe detector demo cfg/combine9k.data cfg/yolo9000.cfg yolo9000.weights

yolo9000 CPU on camera #0

darknet_no_gpu.exe detector demo cfg/combine9k.data cfg/yolo9000.cfg yolo9000.weights

Image Classification

Download weights

https://pjreddie.com/darknet/imagenet/

cfg	weights
cfg/alexnet.cfg	https://pjreddie.com/media/files/alexnet.weights
cfg/vgg-16.cfg	https://pjreddie.com/media/files/vgg-16.weights
cfg/extraction.cfg	https://pjreddie.com/media/files/extraction.weights
cfg/darknet.cfg	https://pjreddie.com/media/files/darknet.weights
cfg/darknet19.cfg	https://pjreddie.com/media/files/darknet19.weights
cfg/darknet19_448.cfg	https://pjreddie.com/media/files/darknet19_448.weights
cfg/resnet50.cfg	https://pjreddie.com/media/files/resnet50.weights
cfg/resnet152.cfg	https://pjreddie.com/media/files/resnet152.weights
cfg/densenet201.cfg	https://pjreddie.com/media/files/densenet201.weights

Train custom darknet19_448 network

0. Fork networks/darknet19-template to networks/my-darknet19-net

1. Download pre-trained weights for the convolutional layers: http://pjreddie.com/media/files/darknet19_448.conv.23 to bin/darknet19_448.conv.23

2. Create file obj.cfg with the same content as in darknet19_448.cfg (or copy darknet19_448.cfg to obj.cfg) and:

• set batch to 128 or 64 or 32 depends on your GPU memory in darknet19-classify.cfg#L4
• change line to subdivisions=4
• set filter-value equal to classes in darknet19-classify.cfg#L189