z1069614715 / ultralytics-face

WIDER-FACE Face Detector Based On YOLOV8

WIDER-FACE Face Detector Based On YOLOV8

BiliBili-Video:深度学习改进实验必看！基于YOLOV8的WIDER-FACE改进(轻量化+提点)实验思路讲解

Environment

1. conda create -n pytorch_2_1_0_py39 python=3.9 anaconda
2. pip install torch==2.1.0 torchvision==0.16.0 --index-url https://download.pytorch.org/whl/cu121
3. pip install -i https://pypi.tuna.tsinghua.edu.cn/simple timm==0.9.16 thop efficientnet_pytorch==0.7.1 einops albumentations==1.4.0 Cython
4. cd widerface_evaluate && python setup.py build install

DataSet Downloads Link

BaiDu Cloud

Contrast Experiment

model	Parameters(M)	GFLOPs	Easy Val AP	Medium Val AP	Hard Val AP
Ours	0.503	5.0	0.937	0.922	0.809
Yolov5n-0.5	0.447	0.571	0.908	0.881	0.738
Yolov5n	1.726	2.111	0.936	0.915	0.805
Yolov7-tiny	-	13.2	0.947	0.926	0.821
SCRFD-10GF	1.62	2.57	0.952	0.939	0.831
SCRFD-2.5GF	0.67	2.53	0.938	0.922	0.785
yolov8n-pose(baseline)	3.08	8.3	0.944	0.919	0.775

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
yolov8n-pose no-pretrain	3,078,128	8.3	6.4m	0.936	0.912	0.776	0.00086s
filter 5 pixel lowprecision object in 640 images-size	3,078,128	8.3	6.4m	0.938	0.917	0.779	0.00086s
filter 5 pixel lowprecision object in 640 images-size + topk=3	3,078,128	8.3	6.4m	0.931	0.915	0.787	0.00086s
filter 5 pixel lowprecision object in 640 images-size + topk=3	3,078,128	8.3	6.4m	0.931	0.915	0.787	0.00086s
filter 5 pixel lowprecision object in 640 images-size + topk=3 + P6	4,878,288	8.3	9.6m	0.935	0.919	0.789	0.00086s
filter 5 pixel lowprecision object in 640 images-size + topk=3 + P6 + HGStem	4,899,840	10.1	9.7m	0.941	0.928	0.810	0.00136s
filter 5 pixel lowprecision object in 640 images-size + topk=3 + P6 + HGStem + LSCD	3,887,489	8.4	7.7m	0.942	0.928	0.807	0.00130s
filter 5 pixel lowprecision object in 640 images-size + topk=3 + P6 + HGStem + LSCD + BIFPN	2,386,962	7.6	4.9m	0.942	0.929	0.811	0.00135s
filter 5 pixel lowprecision object in 640 images-size + topk=3 + P6 + HGStem + LSCD + BIFPN + Rep	2,386,962	7.6	4.9m	0.943	0.929	0.814	0.00135s
filter 5 pixel lowprecision object in 640 images-size + topk=3 + P6 + HGStem + LSCD + BIFPN + Rep + LMAP 1.5X	503,686	5.0	1.3m	0.937	0.922	0.809	0.00111s

Test Image (image-size:1280,conf:0.25,max_det:1000)

Experimental Result BaseLine:yolov8n-pose

yolov8n-pose

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
yolov8n-pose	3,078,128	8.3	6.4m	0.944	0.919	0.775	0.00086s
yolov8n-pose no-pretrain	3,078,128	8.3	6.4m	0.936	0.912	0.776	0.00086s

(yolov8n-pose no-pretrain) filter x pixel lowprecision object in 640 images-size

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
yolov8n-pose	3,078,128	8.3	6.4m	0.936	0.912	0.776	0.00086s
filter 9 pixel lowprecision object in 640 images-size	3,078,128	8.3	6.4m	0.943	0.922	0.758	0.00086s
filter 7 pixel lowprecision object in 640 images-size	3,078,128	8.3	6.4m	0.941	0.918	0.768	0.00086s
filter 6 pixel lowprecision object in 640 images-size	3,078,128	8.3	6.4m	0.939	0.918	0.774	0.00086s
filter 5 pixel lowprecision object in 640 images-size	3,078,128	8.3	6.4m	0.938	0.917	0.779	0.00086s
filter 4 pixel lowprecision object in 640 images-size	3,078,128	8.3	6.4m	0.934	0.911	0.775	0.00086s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size) + FaceRandomCrop

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	3,078,128	8.3	6.4m	0.938	0.917	0.779	0.00086s
FaceRandomCrop(max_crop_ratio=0.5, p=0.2)	3,078,128	8.3	6.4m	0.934	0.909	0.764	0.00086s
FaceRandomCrop(max_crop_ratio=0.2, p=0.5)	3,078,128	8.3	6.4m	0.929	0.902	0.741	0.00086s
FaceRandomCrop(max_crop_ratio=0.2, p=0.2)	3,078,128	8.3	6.4m	0.933	0.909	0.766	0.00086s
FaceRandomCrop(max_crop_ratio=0.1, p=0.2)	3,078,128	8.3	6.4m	0.935	0.910	0.767	0.00086s
FaceRandomCrop(max_crop_ratio=0.1, p=0.1)	3,078,128	8.3	6.4m	0.936	0.915	0.773	0.00086s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size) + TAL

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	3,078,128	8.3	6.4m	0.938	0.917	0.779	0.00086s
topk=7	3,078,128	8.3	6.4m	0.934	0.915	0.782	0.00086s
topk=5	3,078,128	8.3	6.4m	0.935	0.917	0.787	0.00086s
topk=3	3,078,128	8.3	6.4m	0.931	0.915	0.787	0.00086s
topk=3 beta=9.0	3,078,128	8.3	6.4m	0.926	0.907	0.789	0.00086s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL) + SPPF-3

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	3,078,128	8.3	6.4m	0.931	0.915	0.787	0.00086s
SPPF-3	3,078,128	8.3	6.4m	0.932	0.915	0.787	0.00086s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL) + P6

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	3,078,128	8.3	6.4m	0.931	0.915	0.787	0.00086s
P6	4,878,288	8.3	9.6m	0.935	0.919	0.789	0.00086s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL) + P6-C2f

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	3,078,128	8.3	6.4m	0.931	0.915	0.787	0.00086s
P6-C2f	4,863,828	8.2	9.6m	0.933	0.916	0.786	0.00083s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL + P6) + ADown

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	4,878,288	8.3	9.6m	0.935	0.919	0.789	0.00086s
ADown	4,874,960	8.1	9.6m	0.934	0.915	0.770	0.00093s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL + P6) + V7Down

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	4,878,288	8.3	9.6m	0.935	0.919	0.789	0.00086s
V7Down	4,876,512	8.3	9.6m	0.935	0.918	0.789	0.00094s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL + P6) + HGStem

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	4,878,288	8.3	9.6m	0.935	0.919	0.789	0.00086s
HGStem	4,899,840	10.1	9.7m	0.941	0.928	0.810	0.00136s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL + P6) + LSCD

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	4,878,288	8.3	9.6m	0.935	0.919	0.789	0.00086s
LSCD	3,865,937	6.6	7.7m	0.937	0.920	0.789	0.00080s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL + P6 + HGStem) + LSCD

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	4,899,840	10.1	9.7m	0.941	0.928	0.810	0.00136s
LSCD	3,887,489	8.4	7.7m	0.942	0.928	0.807	0.00130s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL + P6 + HGStem + LSCD) + BIFPN

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	3,887,489	8.4	7.7m	0.942	0.928	0.807	0.00130s
BIFPN	2,386,962	7.6	4.9m	0.942	0.929	0.811	0.00135s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL + P6 + HGStem + LSCD + BIFPN) + Rep

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	2,386,962	7.6	4.9m	0.942	0.929	0.811	0.00135s
Rep1	2,386,962	7.6	4.9m	0.943	0.929	0.814	0.00135s
Rep2	2,386,962	7.6	4.9m	0.941	0.925	0.812	0.00135s
Rep3	2,386,962	7.6	4.9m	0.940	0.926	0.814	0.00135s
Rep4	2,386,962	7.6	4.9m	0.941	0.927	0.813	0.00135s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL + P6 + HGStem + LSCD + BIFPN + Rep) + C2f-XXX

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	2,386,962	7.6	4.9m	0.943	0.929	0.814	0.00135s
EMBC	2,623,842	7.1	5.4m	0.940	0.926	0.813	0.00156s
Faster	1,901,922	6.6	4.0m	0.935	0.922	0.808	0.00132s
DWR	2,303,010	7.6	4.8m	0.939	0.924	0.813	0.00137s
RVB	1,886,370	6.6	4.0m	0.938	0.924	0.808	0.00136s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL + P6 + HGStem + LSCD + BIFPN + Rep) + SlideLoss

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	2,386,962	7.6	4.9m	0.943	0.929	0.814	0.00135s
SlideLoss	2,386,962	7.6	4.9m	0.942	0.928	0.814	0.00135s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL + P6 + HGStem + LSCD + BIFPN + Rep) + NWD

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	2,386,962	7.6	4.9m	0.943	0.929	0.814	0.00135s
iou:0.5 nwd:0.5 c:1	2,386,962	7.6	4.9m	0.940	0.925	0.810	0.00135s
iou:0.7 nwd:0.3 c:1	2,386,962	7.6	4.9m	0.941	0.925	0.811	0.00135s
iou:0.5 nwd:0.5 c:24.4	2,386,962	7.6	4.9m	0.942	0.929	0.813	0.00135s
iou:0.7 nwd:0.3 c:24.4	2,386,962	7.6	4.9m	0.941	0.927	0.813	0.00135s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL + P6 + HGStem + LSCD + BIFPN + Rep) + Inner-CIoU

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	2,386,962	7.6	4.9m	0.943	0.929	0.814	0.00135s
ratio=1.1	2,386,962	7.6	4.9m	0.934	0.923	0.810	0.00135s
ratio=1.2	2,386,962	7.6	4.9m	0.942	0.927	0.815	0.00135s
ratio=1.25	2,386,962	7.6	4.9m	0.940	0.926	0.811	0.00135s
ratio=1.3	2,386,962	7.6	4.9m	0.942	0.928	0.814	0.00135s
ratio=1.4	2,386,962	7.6	4.9m	0.941	0.926	0.813	0.00135s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL + P6 + HGStem + LSCD + BIFPN + Rep) + XIoU

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	2,386,962	7.6	4.9m	0.943	0.929	0.814	0.00135s
EIoU	2,386,962	7.6	4.9m	0.939	0.925	0.814	0.00135s
DIoU	2,386,962	7.6	4.9m	0.941	0.926	0.813	0.00135s
SIoU	2,386,962	7.6	4.9m	0.935	0.923	0.814	0.00135s
MODIoU	2,386,962	7.6	4.9m	0.940	0.927	0.813	0.00135s

(yolov8n-pose no-pretrain filter 5 pixel lowprecision object in 640 images-size + TAL + P6 + HGStem + LSCD + BIFPN + Rep) + LAMP

model	Parameters	GFLOPs	Model Size	Easy Val AP	Medium Val AP	Hard Val AP	Inference Time(bs:32)
baseline	2,386,962	7.6	4.9m	0.943	0.929	0.814	0.00135s
LMAP 1.5X	503,686	5.0	1.3m	0.937	0.922	0.809	0.00111s
LMAP 2.0X	321,762	3.7	1.0m	0.919	0.907	0.792	0.00090s
LMAP 2.5X	229,801	2.9	0.8m	0.904	0.892	0.774	0.00077s

Reference

1. https://github.com/deepinsight/insightface/tree/master/detection/scrfd
2. https://github.com/deepcam-cn/yolov5-face
3. https://github.com/derronqi/yolov7-face
4. https://github.com/derronqi/yolov8-face
5. https://github.com/Krasjet-Yu/YOLO-FaceV2