Real Time Object Detection By Using YOLO to online shopping
To training your custom dataset by using Darknet,
1- You must first tag your dataset. You can download draw boxing program from this url https://github.com/puzzledqs/BBox-Label-Tool.
-
After download draw boxing program, create new file like as 002 into folders of "Examples" , "Images" and "Labels".
-
You copy your images in the daatset to a your new file in "Images" file.For this tutorial, path is /Images/002/.
-
If extension of your images are different from "jpeg" ,change all the "jpeg" values in the main.py file to your extension. For example, making "jpeg" to "jpg".
-
Open cmd or terminal and run "python main.py"
-
You see interface of program and than write images folder name in Image Dir textbox .(For this tutorial , it is 002) Click Load button.Then draw box all of your images.
-
You can see text files of your label coordinates.
-
Note: You have to create files each of class!!
-
You have to convert to float value your coordinates values.So, you download convert.py file and place it in the BBox-Label-Tool project folder.
-
Change some values: 1- line of 15 : classes= ["002"]
2- line of 34 : yourPath="your path" (Example my path: "C:/BBox-Label-Tool-Python3.x/Labels/002")
3- line of 35 : ypurOutputPath="your output path"
4- line of 37 : cls="002" -
Than run "python convert.py" at terminal.
2- Download darknet project from https://github.com/AlexeyAB/darknet . To use on Windows you follow darknet->build->darknet->x64
Note-1: You must use cudnn 9.0 version. So, you change "9.1" to "9.0" in darknet.sln to build your project.
Note-2: You must use opencv 3.4.0. and you put "opencv_ffmpeg340_64.dll" and "opencv_world340.dll" files next to darknet.exe file.
-
Create custom dataset folder into path of /x64/data/ . Put your images , label texts and process.py file into this new folder from BBox-Label-Tool project.
-
To create test and trainig dataset : run process.py ( Like as a C:\darknet\build\darknet\x64\data\custom_dataset>python process.py )
3- We prepare dataset to training. Now, we have to create ".data", ".names" and ".cfg" files for custom dataset. Note: There is many yolo architecture. But, I suggest you to use tiny yolo architecture for android. Because of, expecially weights file size of Yolov2 or Yolov3 are too large for android.
1- Create custom.data from any one of "--.data" file and edit according to yourself like this:
2- Create custom.names from one of "--.names" file and edit according to your class names.
3- Create "custom_yolov2_tiny.cfg" from one of "yolov2-tiny.cfg" file. And change this file like this:
1- Lines of 3 and 6 : batch=64
2- Lines of 4 and 7 : subdivisions=32
3- Line of 237 : filters = 30 ---(filters= (count of class + 5)*5)
4- Line of 244 : classes = 1 --- classes= count of class
4- Finally, put "yolov2-tiny.conv.13" file next to darknet.exe file.
5- Start Training: darknet.exe detector train data/custom.data cfg/custom_yolov2_tiny.cfg yolov2-tiny.conv.13
(C:\darknet\build\darknet\x64>darknet.exe detector train data/custom.data cfg/custom_yolov2_tiny.cfg yolov2-tiny.conv.13) You can see the avarage loss value for each iterations and the program save weights file into backup folder for every 100 iterasyon automatically.
6- You can stop trainig when the average loss value is between 0.8 and 1.0. 7- To use weights file for android, you have to transformation to protobuf file.
1- Download darkflow from https://github.com/thtrieu/darkflow.
2- Create bin folder into darkflow project.
3- Put your weights file into bin forder.
4 Put your cfg file into darkflow/cfg folder.
5- Convert pb file like this:
C:\darkflow> python flow --model cfg/custom_yolov2_tiny.cfg --load bin/custom_yolov2_tiny_3500.weights --savepb
8- Now you have probuf file.(darkflow/built_graph/) 9- You can add your own protobuf file to the asset file in android and you can run the program.
Finally, the working version of the application can be found at this link: https://www.youtube.com/watch?v=q8Ka7dzFymE&feature=youtu.be