River-Cold / slim-neck-by-gsconv

Pytorch implementation of the 'Slim-neck by GSConv: A better design paradigm of detector architectures for autonomous vehicle'

Slim-neck by GSConv: A better design paradigm of detector architectures for autonomous vehicles

paper

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

Model	size (pixels)	mAPval 0.5:0.95	mAPval 0.5	FPS T4 b1	FPS T4 b32	params (M)	FLOPs @640 (G)
yolov5n(ultralytics)	640	28.0	45.7	--	--	1.9	4.5
GSyolov5n	640	28.4(+0.4)	47.0(+1.3)	147	207	1.8	4.0
Model	size (pixels)	mAPval 0.5:0.95	mAPval 0.5	FPS A40 b1	FPS A40 b32	params (M)	FLOPs @640 (G)
yolov5s	640	35.7	54.3	109	297	7.2	16.4
GSyolov5s	640	36.0(+0.3)	54.2	95	312(+15)	7.0	14.5

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 better design paradigm of detector architectures for autonomous vehicles},
author={Li, Hulin and Li, Jun and Wei, Hanbing and Liu, Zheng and Zhan, Zhenfei and Ren, Qiliang},
journal={arXiv preprint arXiv:2206.02424},
year={2022}
}

基于GSConv的轻量融合层：一个更好的轻量化检测器结构设计范式用于自动驾驶

论文

实验数据集:
- PASCAL VOC 2007+12 (通用检测器)
- WiderPerson (用于行人检测)
- SODA10M (用于自动驾驶)
- DOTA1.0 (用于遥感影像目标检测)
(我们只提供我们所使用的训练、验证和测试的文本文档,以便于您复现我们的结果。 被标注的图像和它们的标注文件请您访问相应的数据集官方网站获取。)

---

遥感小目标检测结果的一个对比例子

scaled-yolov4

slim neck scaled-yolov4

---

训练自定义数据集

1. 训练基于GSConv的轻量化yolov5检测器

(7月14日更新)

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. 训练基于GSConv的轻量化scaled-yolov4检测器

(8月17日更新)

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

预训练权重

MS COCO

Model	size (pixels)	mAPval 0.5:0.95	mAPval 0.5	FPS T4 b1	FPS T4 b32	params (M)	FLOPs @640 (G)
yolov5n(ultralytics)	640	28.0	45.7	--	--	1.9	4.5
GSyolov5n	640	28.4(+0.4)	47.0(+1.3)	147	207	1.8	4.0
Model	size (pixels)	mAPval 0.5:0.95	mAPval 0.5	FPS A40 b1	FPS A40 b32	params (M)	FLOPs @640 (G)
yolov5s	640	35.7	54.3	109	297	7.2	16.4
GSyolov5s	640	36.0(+0.3)	54.2	95	312(+15)	7.0	14.5

验证和测试基于GSConv的轻量检测器的性能

1. 测试基于GSConv的轻量化yolov5检测器

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

2. 测试基于GSConv的轻量化scaled-yolov4检测器

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

参考

• 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

引用此工作

@article{li2022slim,
title={Slim-neck by GSConv: A better design paradigm of detector architectures for autonomous vehicles},
author={Li, Hulin and Li, Jun and Wei, Hanbing and Liu, Zheng and Zhan, Zhenfei and Ren, Qiliang},
journal={arXiv preprint arXiv:2206.02424},
year={2022}
}