Masked Face Detection

Summary

Backbone: Shufflenet v2 + FPN

Prediction Head: YoloV5

Benchmark

Model Size mAP (COCO) FLOPS (G) Params (M) Jetson Nano (FPS)
Shufflnet v2-yolo 352x352 24.1 0.22 0.24 33*
YoloV5n 640x640 / 17 7.3 4-5

GIF 2022-1-20 21-36-41.gif

Dataset

Face Mask Label Dataset (FMLD) + Kaggle Medical Mask Dataset

Label size
masked 27,985
unmasked 25,267
incorrectly 1,330
TOTAL 54,582

chart.png

Samples

test00_result.jpg

180_mask_result.jpg

test_00001973_result.jpg

test_00003519_result.jpg

Precision Recall [email protected] F1
0.76 0.55 0.55 0.61

YoloV5

https://github.com/pritul2/yolov5_FaceMask/blob/master/yolov5_train.ipynb

download.png

https://github.com/offsouza/yolov5_face_mask_detection

68747470733a2f2f6d69726f2e6d656469756d2e636f6d2f6d61782f313238302f312a76374b795f57795f454f7146656251345176662d44412e706e67.png

Todo (optimization)