JamesYang568 / F-YOLO

Code of 'F-YOLO: Delving into Fuzzy YOLO for Improved Traffic Object Detection'

F-YOLO

Code of paper F-YOLO: Delving into Fuzzy YOLO for Improved Traffic Object Detection

Fuzzy Fusion Module

FFM is deployed to steer the fusion process, managing feature uncertainty during the fusion while achieving multi-scale feature fusion.

Fuzzy Attention

The fuzzy attention mechanism consists of two distinct submodules: CFA and SFA. The former targets the mitigation and control of spatial dimension uncertainty, while the latter addresses similar concerns in the channel dimension.

CFA

Channel fuzzy attention employs fuzzy methodologies to evaluate each channel's contribution to the detection task.

SFA

Spatial fuzzy attention focuses on the uncertainty of each position in feature space, rather than their values.

GPB

GPB is designed to more effectively assimilate global context information, which is composed by GhostBlock and Pooling Enhancement submodule.

Installation

1. Pytorch==1.2.0
2. MMDetection==3.0.0, MMEngine==0.7.3
3. MMYOLO==0.5.0

For MMYOLO, please see https://mmyolo.readthedocs.io/

Placing files in our repository to the appropriate place as in MMYOLO, just overwrite or copy the new files is enough.

Caution: make sure all files is downloaded and renewed, especially __init__.pys, mmyolo/tools/analysis_tools and mmyolo/mmyolo/models/layers

Usage

Same as MMYOLO, please see 15 minutes to get started with MMYOLO object detection — MMYOLO 0.5.0 documentation and our config files are in mmyolo/configs/yolov6s/.

We used a very simple and intuitive way to present the code, with all modules plug and play.

train

bash ./tool/dist_train.sh 2

inference

# for val&test
python ./tool/test.py

benchmark

python ./tool/analyze_logs/benchmark.py
python ./tool/analyze_logs/analyze_logs.py
python ./tool/analyze_logs/get_flops.py

Result

TABLE I. Experimental results on COCO val2017

Method	Backbone	AP	AP50	AP75	APS	APM	APL
PPYOLOEs	CSPRepResNet	43.1	60.5	46.6	23.2	46.4	56.9
YOLOv6s	EfficientRep	43.5	60.4	46.8	23.7	48.9	59.9
RTMDets	Modified CSP	44.5	61.6	48.3	24.1	49.2	61.9
YOLOv8s	Modified CSP	44.2	61.1	47.9	25.9	49.1	60.1
DAMO-YOLOs	MAE-NAS	46.0	61.9	49.5	25.9	50.6	62.5
YOLO-MSs	MS-Block	46.2	63.7	50.5	26.9	50.5	63.0
Ours	EfficientRep	46.2	65.5	51.0	25.9	52.4	64.4

TABLE II. Experimental results on VisDrone2017

Method	Backbone	AP	AP50	AP75	APS	APM	APL
YOLOXs	Modified CSP	17.6	33.9	16.2	9.0	27.7	45.0
YOLOv6s	EfficientRep	19.5	33.2	19.5	9.7	30.8	47.4
PPYOLOEs	CSPRepResNet	20.0	34.6	20.0	10.5	31.5	51.0
UAV-YOLO	Modified CSP(YOLOv5)	20.6	37.6	-	-	-	-
YOLOv8s	Modified CSP(C2f)	20.9	36.8	20.7	10.7	33.2	49.7
FE-YOLOv5	Modified CSP(YOLOv5)	21.0	37.0	20.7	13.2	29.5	39.1
Ours	EfficientRep(YOLOv6)	23.4	38.2	24.2	11.3	37.6	59.1

TABLE III. Experimental results on UA-DETRAC

Method	Backbone	AP	AP50	AP75	APS	APM	APL	FLOPs
FasterRCNN	ResNet	38.8	66.3	40.3	3.9	28.9	43.7	60.5
RetinaNet	ResNet	32.1	56.6	33.8	2.1	22.8	46.8	45.2
CascadeRCNN	ResNet	41.1	65.2	46.4	3.8	27.5	57.1	88.1
ATSS	ResNet	32.1	58.8	31.0	4.1	24.0	45.1	49.2
YOLOv5s	Modified CSP	44.9	62.0	51.8	5.2	34.7	56.2	17.1
YOLOXs	Modified CSP	40.0	56.2	46.6	3.4	28.8	49.6	26.8
PPYOLOEs	CSPRepResNet	43.7	55.3	51.1	1.1	28.8	53.5	17.4
YOLOv6s	EfficientRep	43.0	59.0	49.9	4.2	29.7	58.7	43.8
YOLO7tiny	Extended-ELAN	43.4	62.8	50.6	4.1	32.6	57.2	13.1
YOLOv8s	Modified CSP(C2f)	46.9	62.6	52.9	3.9	31.4	58.7	28.6
Ours	EfficientRep	46.7	61.1	54.9	9.0	27.4	59.2	44.3

TABLE IV. mAP of different classes in BDD100K

Models	Ped	Rider	Car	Truck	Bus	Train	Bicycle	TL	TS
PPYOLOEs	55.3	37.2	76.3	61.4	60.4	0.0	42.9	57.6	66.1
YOLOv6s	59.6	42.0	77.4	61.8	59.6	0.0	40.6	58.3	61.9
YOLOv8s	62.2	45.9	79.8	64.0	61.4	0.6	44.0	62.3	65.4
Ours	62.9	46.2	79.4	64.1	62.0	7.0	44.5	61.9	65.7

TABLE V. Ablation study

FA	FFM	GPB	AP	AP50	AP75	APS	APM	APL
			43.5	60.4	46.8	23.7	48.9	59.9
		√	44.0	61.9	48.2	23.6	49.2	61.0
	√		44.9	62.6	48.4	24.6	49.8	62.5
√			44.5	62.3	47.9	24.4	49.1	62.8
√	√		45.7	63.5	49.1	25.4	50.3	65.0
√	√	√	46.2	65.5	51.0	25.9	52.4	64.4

①

The F-YOLO model are trained from scratch on the MS COCO and UA-DETRAC datasets for 300 epochs, without relying on any baseline model checkpoints. Subsequent fine-tuning was carried out on the VisDrone2019 and BDD100K datasets for 50 and 100 epochs, respectively.

②

Train and test on 2 NVIDIA GeForce RTX 3090 GPUs.

Models

Pretrained checkpoint will coming soon.

model	epoch	AP	config file	download
F-YOLO_coco	300	46.2	cfg	[link]
F-YOLO_visdrone	100	23.4	cfg	[link]
F-YOLO_bdd	100	32.2	cfg	[link]
F-YOLO_ua-detrac	300	46.7	cfg	[link]
yolox_ua-detrca	300	40.0	cfg	[link]
PPYOLOE_ua-detrac	100	43.7	cfg	[link]
PPYOLOE_bdd	100	29.2	cfg	[link]
YOLOv6_coco	300	43.5	cfg	link code:hwpt
YOLOv6_bdd	100	29.7	cfg	[link]
YOLOv6_ua-detrac	300	43.0	cfg	[link]
YOLOv8_visdrone	100	20.9	cfg	[link]
YOLOv8_bdd	100	31.5	cfg	[link]

Contact

jiaxiongyang at tongji dot edu dot cn