Record the use of yolov5 to detect rotating targets

Because the internship company needs to use the detection of rotating targets , So I learned relevant knowledge these days , And found a lot of information , Many detours have been spared . The whole implementation process of the project is recorded below .

I refer to the following bloggers ：

DOTAv2 Remote sensing image rotating target detection competition experience sharing （Swin Transformer + Anchor free/based programme ） - You know

The personal space of chicken stewing technology _ Bili, Bili _Bilibili 

First , First create a virtual environment for the project

VS2015

pytorch 1.6.0

torchvision 0.7.0

Among them the VS2015,17,19 All versions are available , however pytorch and torchvision Must be the required version  , Otherwise, it could go wrong . Proposed installation pytorch and torchvision The offline installation package is used for installation , More stable . Address ,

https://download.pytorch.org/whl/torch_stable.html

Then open Anaconda prompt Create a virtual environment for the project

conda create -n royolov5 python=3.8
conda activate yolov5

After entering the created virtual environment cd Go to download pytorch and torchvision Folder . And then execute

pip install torch-1.6.0+cu101-cp38-cp38-win_amd64.whl

pip install torchvision-0.7.0+cu101-cp38-cp38-win_amd64.whl

Note the suffix here .whl Don't forget to .

Come here , will pytorch The frame is set up .

next , Download the compressed package I sent , After decompressing , Again cd Get into YOLOv5_DOTA_OBB-master Home directory of project files

perform  pip install -r requirements.txt  Install the relevant package .

Again , add to E:\PytorchPro\YOLOv5_DOTA_OBB-master\swigwin-4.0.2 To path Inside , Here you need to refresh the system environment , You can choose to restart , You can also enter cmd, Input set path=c refresh

Again , stay Anaconda prompt Medium royolov5 In the environment cd Get into utils Run respectively

swig -c++ -python polyiou.i

python setup.py build_ext --inplace

Again , take E:\ProgramData\Anaconda3\envs\royolov5\Lib\site-packages\shapely\DLLs all dll Copied to the E:\ProgramData\Anaconda3\envs\royolov5\Library\bin

To this step , The training environment is basically ready , Now let's start preparing the data set required for the rotation target

First, use rolabelimg Rotate target dimension tool , How to use self Baidu . Download address

https://github.com/cgvict/roLabelImg

After labeling the image, you get xml Format , Here you need to convert the data format , accord with yolov5 Of txt Format

First , function roxml_to_data.py, Quasi switch to dota Of 8 Point data format

Again , Get into DOTA_devkit_YOLO-master, modify DOTA_devkit_YOLO-master\dota_utils.py Medium classnames_v1_5, Modify the type of data marked for you

Again , Run respectively PaddingPIC.py and YOLO_Transform.py

among ,PaddingPIC.py The function of is to expand the picture to the same height and width , Easy data processing .YOLO_Transform.py The role of the dota The data format of is converted to yolov5 Data format .

Again , take yolov5 Pictures and... For training txt Put the files into DOTA_demo_view In a folder images and labels in

 

Revise data\DOTA_ROTATED.yaml in nc and names, Or write it yourself yaml file

You can modify the network model yaml file , If not, use the official yaml Just file

Finally, don't forget to download the weight file , Pay attention here , Because this project is based on v5-3.1 Version of , therefore , The latest weight file cannot be downloaded pt file , Otherwise, it will report a mistake .

Last , Just start training .