CenterNet

Introduction

Network for CenterNet. The pytorch implementation for "Objects as Points ".

Quick start

1. Clone this repository

git clone https://github.com/Runist/torch_CenterNet

2. Install torch_CenterNet from source.

cd torch_CenterNet
pip install -r requirements.txt

3. Download the Pascal dataset or COCO dataset. Create new folder name call "data" and symbolic link for your dataset.

mkdir data
cd data
ln -s xxx/VOCdevkit VOCdevkit
cd ..

4. Prepare the classes information file and place it in "data" directory, the txt file format is:

aeroplane
bicycle
...
tvmonitor

5. Configure the parameters in tools/args.py.
6. Start train your model.

python tools/train.py

or use Linux shell to start.

sh scripts/train_yolo.sh

7. Open tensorboard to watch loss, learning rate etc. You can also see training process and training process and validation prediction.

tensorboard --logdir ./weights/yolo_voc/log/summary

8. After train, you can run evaluate.py to watch model performance.

python tools/evaluate.py

As well as use Linux shell to start.

sh scripts/eval_yolo.sh

9. Get prediction of model.

python tools/predict.py

Or use script to run

sh scripts/predict.sh

Train your dataset

We provide three dataset format for this repository "yolo", "coco", "voc",You need create new annotation file for "yolo", the format of "yolo" is:

image_path|1,95,240,336,19
image_path|305,131,318,151,14|304,142,354,160,3

"coco", "voc" is follow the format of their dataset. And prepare the classes information file and place it in "data" directory.

Performance

Train Dataset	Val Dataset	weight	mAP 0.5	mAP 0.5：0.95
VOC07+12	VOC-Test07	resnet50-CenterNet.pt	0.622	0.436

Reference

Appreciate the work from the following repositories:

• bubbliiiing/centernet-pytorch

• katsura-jp/pytorch-cosine-annealing-with-warmup

• YunYang1994/tensorflow-yolov3

License

Code and datasets are released for non-commercial and research purposes only. For commercial purposes, please contact the authors.