Skip to content

Latest commit

 

History

History
126 lines (113 loc) · 9.23 KB

README.md

File metadata and controls

126 lines (113 loc) · 9.23 KB
Raw

SFCHD-SCALE

The dataset and code of the paper "Large, Complex, and Realistic Safety Clothing and Helmet Detection: Dataset and Method".
Authors: Fusheng Yu†, Jiang Li†, Xiaoping Wang, Shaojin Wu, Junjie Zhang, Zhigang Zeng († Equal Contribution)
Affiliation: Huazhong University of Science and Technology (HUST)

Citation

@article{yu2024sfchd-scale,
  title={Large, Complex, and Realistic Safety Clothing and Helmet Detection: Dataset and Method},
  author={Fusheng Yu and Jiang Li and Xiaoping Wang and Shaojin Wu and Junjie Zhang and Zhigang Zeng},
  year={2024},
  journal = {},
  volume = {},
  pages = {},
  doi={}
}

Dataset Information

Comparison between SFCHD and existing open-source datasets for safety helmets

Dataset Year #Category #Sample #Instance Color Task Data Source
Pictor-v3 2020 6 1,330 9,208 RGB Detection Web-mined and Crowd-sourced
SHWD 2019 2 3,241 10,457 RGB Detection Web-mined
SFCHD 2023 7 12,373 50,558 RGB Detection Chemical Plant

Statistics of instance distribution per category in the SFCHD dataset

Category Person Safety Helmet Safety Clothing Other Clothing Head Blurred Clothing Blurred Head Total
Training 13,528 11,378 11,781 626 961 1,053 896 40,223
Testing 3,482 2,920 3,032 154 239 271 238 10,336
Total 17,010 14,298 14,813 780 1,200 1,324 1,134 50,559

Category distribution of objects with different sizes in the SFCHD dataset

Category Total Large Medium Small
Safety Helmet 14,298 3,551 3,030 7,717
Head 1,200 95 164 941
Blurred Clothing 1,324 281 313 730
Blurred Head 1,134 15 41 1,078

Experimental Results

Comparisons of different methods on the Pictor-v3, SHWD, and SFCHD datasets [mAP(0.50)/mAP(0.50:0.95)]

Method Backbone Pictor-v3 SHWD SFCHD (ours)
SSD VGG16 85.5 / 48.8 80.8 / 57.4 72.8 / 41.5
Faster RCNN ResNet-50 90.6 / 53.4 84.8 / 63.1 76.4 / 50.3
FCOS ResNet-50 89.5 / 52.4 85.8 / 63.9 76.4 / 49.6
VFNet ResNet-50 91.4 / 55.2 85.7 / 63.9 76.4 / 51.0
RetinaNet ResNet-50 90.5 / 54.4 85.5 / 63.6 75.9 / 48.9
TOOD ResNet-50 91.5 / 55.8 86.7 / 64.4 78.9 / 52.3
YOLOv5 CSPDarknet53 88.2 / 53.6 84.0 / 63.9 74.1 / 49.6

Performance of different categories in the SFCHD dataset

Method Person Safety Helmet Safety Clothing Other Clothing Head Blurred Clothing Blurred Head mAP(0.50:0.95) mAP(0.50)
SSD 60.2 56.5 55.7 45.2 38.4 20.5 14.2 41.5 72.8
Faster RCNN 71.2 64.7 64.6 54.5 49.3 27.2 20.5 50.3 76.4
FCOS 68.4 63.4 64.6 54.0 48.0 28.9 19.6 49.6 76.4
VFNet 73.1 66.2 64.3 54.0 52.5 25.1 22.0 51.0 76.4
RetinaNet 71.1 63.5 64.5 51.8 48.5 27.1 15.9 48.9 75.9
TOOD 72.9 66.0 65.9 56.2 52.8 29.6 22.3 52.3 78.9
YOLOv5 72.7 66.4 63.7 54.9 50.7 21.2 18.9 49.6 74.1

Performance comparisons between our SCALE-YOLO and existing models on the ExDark dataset

Method Bicycle Boat Bottle Bus Car Cat Chair Cup Dog Motorbike People Table mAP(0.50)
YOLOv3 79.8 75.3 78.1 92.3 83.0 68.0 69.0 79.0 78.0 77.3 81.5 55.5 76.4
KinD 80.1 77.7 77.2 93.8 83.9 66.9 68.7 77.4 79.3 75.3 80.9 53.8 76.3
MBLLEN 82.0 77.3 76.5 91.3 84.0 67.6 69.1 77.6 80.4 75.6 81.9 58.6 76.8
Zero-DCE 84.1 77.6 78.3 93.1 83.7 70.3 69.8 77.6 77.4 76.3 81.0 53.6 76.9
MAET 83.1 78.5 75.6 92.9 83.1 73.4 71.3 79.0 79.8 77.2 81.1 57.0 77.7
DENet 80.4 79.7 77.9 91.2 82.7 72.8 69.9 80.1 77.2 76.7 82.0 57.2 77.3
IAT-YOLO 79.8 76.9 78.6 92.5 83.8 73.6 72.4 78.6 79.0 79.0 81.1 57.7 77.8
PE-YOLO 84.7 79.2 79.3 92.5 83.9 71.5 71.7 79.7 79.7 77.3 81.8 55.3 78.0
SCALE-YOLO 81.3 79.3 78.2 93.9 84.2 75.5 74.9 82.3 81.0 77.5 82.5 57.3 79.0

Performance improvements of the SCALE module on the ExDark dataset

Method Bicycle Boat Bottle Bus Car Cat Chair Cup Dog Motorbike People Table mAP(0.50)
FCOS 75.5 64.4 68.0 86.8 78.5 69.3 55.4 71.7 70.0 64.8 72.3 46.7 68.6
FCOS+SCALE 75.1 66.6 73.5 89.9 78.9 67.0 57.2 72.8 74.2 67.3 72.0 45.9 70.0
VFNet 77.4 70.5 76.6 90.6 81.8 67.1 59.4 71.8 72.6 70.6 77.7 53.3 72.5
VFNet+SCALE 79.4 70.2 76.5 89.7 81.7 71.9 60.9 71.5 75.0 71.2 77.4 55.5 73.4
TOOD 77.0 69.2 72.2 90.0 80.0 72.6 63.0 71.8 71.0 71.9 76.2 52.2 72.3
TOOD+SCALE 77.9 70.0 78.3 90.0 80.7 69.1 62.0 72.4 73.7 69.2 78.1 54.2 73.0

Ablation analysis for different pathways in our SCALE module

Method SAP CAP mAP(0.50)
YOLOv3 -- -- 76.4
SCALE-YOLO (Ours) ✓ ✗ 77.3
SCALE-YOLO (Ours) ✗ ✓ 77.8
SCALE-YOLO (Ours) ✓ ✓ 79.0

Performance improvements of the SCALE module on the SFCHD dataset

Method Person Safety Helmet Safety Clothing Other Clothing Head Blurred Clothing Blurred Head mAP(0.50:0.95) mAP(0.50)
FCOS 68.4 63.4 64.6 54.0 48.0 28.9 19.6 49.6 76.4
FCOS+SCALE 68.5 63.8 64.7 53.0 47.6 28.4 20.5 49.5 76.3
VFNet 73.1 66.2 64.3 54.0 52.5 25.1 22.0 51.0 76.4
VFNet+SCALE 73.2 66.5 64.4 53.9 52.1 25.8 23.7 51.4 76.6
TOOD 72.9 66.0 65.9 56.2 52.8 29.6 22.3 52.3 78.9
TOOD+SCALE 72.9 66.2 66.2 56.2 51.6 29.6 23.5 52.4 79.3
YOLOv8 73.6 64.9 67.5 58.5 45.5 32.3 23.9 52.2 77.9
YOLOv8+SCALE 74.4 66.1 68.8 58.4 47.5 33.2 25.2 53.3 78.6

Dataset Acquisition

.
├── annotations
├── classes.txt
├── directory.md
├── images
├── labels
├── labels.cache
├── new_split_yolo
├── sd_train
├── train
├── Vision
└── yolo
8 directories, 3 files

Download the dataset from [链接：https://pan.baidu.com/s/1k2pWg8r-G3KSI2Q3Tdt6kg 提取码：v4ao], unzip and copy the files from images into dataset_SFCHD/images. Also, unzip labels.zip and yolo.zip.

Google Drive: https://drive.google.com/file/d/1-2z7r3J4sZdLvVt5mllvSEwAFO49Y-zj/view?usp=sharing