CenterNet-pytorch-detection-simple-tutorial

A simple demo to study the CenterNet

environment

tqdm>=4.32.2
numpy>=1.15.0
torch>=1.0
torchvision>=0.2.1
Pillow>=7.1.1

data

SeaShips dataset：7000 images，1920x1080 size，6 claases http://www.lmars.whu.edu.cn/prof_web/shaozhenfeng/datasets/SeaShips%287000%29.zip

data preprocess

download the SeaShips dataset and unzip

git clone https://github.com/JavisPeng/CenterNet-pytorch-detection-simple-tutorial.git
cd CenterNet-pytorch-detection-simple-tutorial/
mkdir VOC2007 & cd VOC2007
wget http://www.lmars.whu.edu.cn/prof_web/shaozhenfeng/datasets/SeaShips%287000%29.zip
unzip SeaShips(7000).zip
python voc.py

after run the commands above, you will get a Serialized outfile "data.pth",it is a dict={'classes_name': classes_name, 'train': train, 'val': val}, and the data format of train or val just like [(img_path,[(bbox1,cls1),(bbox2,cls2),])..]

train

train a model from scratch

python main.py train

test

predict a image, w.pth is a trained model weight file

python main.py test --ckpt w.pth --test_img_path VOC2007/JPEGImages/000001.jpg

an output demo

reference

https://github.com/xingyizhou/CenterNet

More detail in my blog https://blog.csdn.net/jiangpeng59/article/details/105732166