yilifan / centernext

The official implementation of "CenterNeXt: Revisiting CenterNet in 2022"

CenterNeXt

The official implementation of "CenterNeXt: Revisiting CenterNet in 2022"

It has not been published yet.

Results and Pre-trained Models

VOC2007

backbone	resolution	mAP	FPS(on Titan Xp)	FPS(on RTX 3090)	FLOPs (G)	model config	weights
ResNet-18	512 x 512	74.92	110	164	14.8	config	model
+ Coupled head	512 x 512	74.61	112	167	14.2	config	model
+ Detection on lower resolution	512 x 512	74.36	127	191	13.0	config	model
+ Mosaic augmentation	512 x 512	74.20	127	191	13.0	config	model
+ Mixup augmentation	512 x 512	75.84	127	191	13.0	config	model
ResNet-50	512 x 512	80.46	65	104	25.0	config	model
ResNet-101	512 x 512	83.00	39	60	44.5	config	model
ConvNeXt-T	512 x 512	83.57	43	87	26.8	config	model

Setup

Create a new conda virtual environment

conda create -n centernext python=3.8 -y
conda activate centernext

Install Pytorch and torchvision following official instructions. For example:

pip install torch==1.10.2+cu113 torchvision==0.11.3+cu113 torchaudio===0.10.2+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

Clone this repo and install required packages:

git clone https://github.com/MarkAny-Vision-AI/CenterNeXt
cd CenterNeXt
pip install -r requirements.txt

Download a dataset:

if (your_os == 'Window'):

cd CenterNeXt
scripts/download-voc0712.bat
python create_gt_files.py

else:

cd CenterNeXt
scripts/download-voc0712.sh
python create_gt_files.py

Training

python train.py --model ./configs/models/your_model.yaml 

On Multi GPU

torchrun --nproc_per_node={number of gpus} train.py --model ./configs/models/your_model.yaml 

Evaluation

python eval.py --model ./configs/models/your_model.yaml --weights /path/to/your_model.pth

Profiling

python profile.py --model ./configs/models/your_model.yaml

License

This project is licensed under the terms of the Attribution-NonCommercial 4.0 International license. It is released for academic research only and is free to researchers from educational or research institutes for non-commercial purposes.

Please see the LICENSE file for more information.

Please contact [email protected] or [email protected] for business inquiries.