Slim-neck by GSConv: a lightweight-design for real-time detector architectures

English | 简体中文

Datasets:
- PASCAL VOC 2007+12
- WiderPerson
- SODA10M (for autonomous vehicles)
- DOTA1.0
(We only provide the train/val/test.txt file we used so that you can reproduce our results. The images & labels can be found on the official websites of these datasets.)

An example of comparison on remote sensing images

scaled-yolov4

slim neck scaled-yolov4

Training the custom datasets

1. For GSConv-yolov5

(Updated July 14th)

git clone https://github.com/AlanLi1997/slim-neck-by-gsconv.git
cd slim-neck-by-gsconv/gsconv-yolov5
pip install requirements.txt
python train.py --cfg models/sm-yolov5s.yaml

2. For GSConv-scaled_yolov4

(Updated Aug 17th)

git clone https://github.com/AlanLi1997/slim-neck-by-gsconv.git
cd slim-neck-by-gsconv
pip install requirements.txt
cd gsconv-scaled-yolov4
python train.py --cfg models/sm-yolov4-p5.yaml

Pretrained Checkpoints

MS COCO

Testing the slim-neck detectors

1. For GSConv-yolov5

cd gsconv-yolov5
python val.py --data yourdata.yaml --weights sm-yolov5s.pt --task test

2. For GSConv-scaled-yolov4

cd gsconv-scaled-yolov4
python val.py --data yourdata.yaml --weights sm-yolov4-p5.pt --task test

References

• https://github.com/ultralytics/yolov5
• https://github.com/AlexeyAB/darknet/tree/yolov4
• https://github.com/WongKinYiu/PyTorch_YOLOv4
• https://github.com/huawei-noah/CV-backbones/tree/master/ghostnet_pytorch
• https://github.com/d-li14/mobilenetv3.pytorch
• https://github.com/megvii-model/ShuffleNet-Series

Citation

@article{li2022slim,
title={Slim-neck by GSConv: a lightweight-design for real-time detector architectures},
author={Li, Hulin and Li, Jun and Wei, Hanbing and Liu, Zheng and Zhan, Zhenfei and Ren, Qiliang},
journal={Real-Time Image Proc 21, 62. https://doi.org/10.1007/s11554-024-01436-6},
year={2024}
}