yanlongbinluck / stronger-centernet

A stronger centernet-based object detection method.

Stronger-CenterNet

This project is the official implementation of the paper 'Scale-balanced real-time object detection with varying input-image resolution' on IEEE Transactions on Circuits and Systems for Video Technology.

Main libs

mmcv==0.2.10
torch==1.1.0
torchvision==0.3.0
cuda==10.1

Install

cd stronger-centernet
python setup.py develop

Training

download the COCO2017 dataset and set the folders as follows:

stronger-centerNet
----data
--------coco
------------annotations
----------------instances_train2017.json
----------------instances_val2017.json
------------train2017
----------------000000169766.jpg
----------------...
------------val2017
----------------000000581781.jpg
----------------...

training with 8 GPUs:

./tools/dist_train.sh ./configs/stronger_centernet/stronger_centernet_resnet18_10x_8GPU.py 8

Evaluation

./tools/dist_test.sh ./configs/stronger_centernet/stronger_centernet_resnet18_10x_8GPU.py ./work_dirs/stronger_centernet_resnet18_10x/stronger_centernet_resnet18_10x_c4518ea4.pth 1

Main results

The FPS is measured on V100 GPU with batchsize = 1, float32 mode. ^+ means model with AFFM and DDH.

backbone	training size	test size	AP@[0.5,...,0.95]	FPS
ResNet-18	768x768	768x768	39.7	90.5
ResNet-18^+	768x768	768x768	41.0	55
ResNet-50	768x768	768x768	43.2	46
ResNet-50^+	768x768	768x768	44.5	38
Darknet-53	768x768	768x768	44.7	42.6
Darknet-53^+	768x768	768x768	45.6	35.8

Acknowledgement

This project is mainly implemented based on ttfnet, mmdetection, CenterNet, etc. Many Thanks for these repos.

Citations

If you use our work in your researches, please cite our paper as follow:

@article{yan2022scale,
  title={Scale-balanced real-time object detection with varying input-image resolution},
  author={Yan, Longbin and Qin, Yunxiao and Chen, Jie},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  volume={33},
  number={1},
  pages={242--256},
  year={2022},
  publisher={IEEE}
}