R-CenterNet(中文)

基于CenterNet的旋转目标检测

前言

本工作初衷是提供一个极其精简的CenterNet代码，并对旋转目标进行检测，1.0为：

 ${R-CenterNet_ROOT}
 |-- backbone
 `-- |-- dlanet.py
     |-- dlanet_dcn.py
 |-- Loss.py
 |-- dataset.py
 |-- train.py
 |-- predict.py

应读者需求，随后更新了2.0

 ${R-CenterNet_ROOT}
 |-- labelGenerator
 `-- |-- Annotations
     |-- voc2coco.py
 |-- evaluation.py

2.0以及data/airplane、imgs、ret文件夹都不是必须的，如果您只是想快速上手，1.0足够了。

demo

• R-DLADCN(推荐)(DCN编译与原版CenterNet保持一致)
  • 
• R-ResDCN(主干网用的ResNet而不是DLA)
  • 
• R-DLANet(如果你不会编译DCN，就使用这个没有编译DCN的主干网)
  • 
• DLADCN.jpg
  • 

常见问题

• 我对CenterNet原版代码 进行了重构，使代码看起来更加简洁。
• 如何编译DCN以及环境需求, 与CenterNet 原版保持一致。
• 关于数据处理与更多细节, 可以参考 here
• torch版本1.2，如果你用的0.4会发生报错。

训练自己的多分类网络

• 打标签用labelGenerator文件夹里面的代码。
• 修改代码中所有num_classes为你的类别数目，并且修改back_bone中hm的数目为你的类别数，如： def DlaNet(num_layers=34, heads = {'hm': your classes num, 'wh': 2, 'ang':1, 'reg': 2}, head_conv=256, plot=False):

Related projects

• R-CenterNet_onnxruntime【C++】
• CenterNet

R-CenterNet(English)

detector for rotated-object based on CenterNet

preface

The original intention of this work is to provide a extremely compact code of CenterNet and detect rotating targets: 1.0

 ${R-CenterNet_ROOT}
 |-- backbone
 `-- |-- dlanet.py
     |-- dlanet_dcn.py
 |-- Loss.py
 |-- dataset.py
 |-- train.py
 |-- predict.py

At the request of readers, 2.0 was subsequently updated：2.0

 ${R-CenterNet_ROOT}
 |-- labelGenerator
 `-- |-- Annotations
     |-- voc2coco.py
 |-- evaluation.py

2.0 and the data/airplane, imgs, ret folders are not required. If you just want to get started quickly, 1.0 is enough。

demo

• R-DLADCN(this code)(How to complie dcn refer to the original code of CenterNet)
  • 
• R-ResDCN(just replace cnn in resnet with dcn)
  • 
• R-DLANet(not use dcn if you don't know how to complie dcn)
  • 
• DLADCN.jpg
  • 

notes

• I refactored the original code to make codes more concise.
• How to complie dcn and configure the environment, refer to the original code of CenterNet.
• For data processing and more details, refer to here
• torch version==1.2，don't use version==0.4!

train your data

• label your data use labelGenerator;
• modify all num_classes to your classes num, and modify the num of hm in your back_bone, such as: def DlaNet(num_layers=34, heads = {'hm': your classes num, 'wh': 2, 'ang':1, 'reg': 2}, head_conv=256, plot=False):

Related projects

• R-CenterNet_onnxruntime【C++】
• CenterNet