All models are trained on WiderFace dataset and weights are from official repositories.
Method | Model | Easy | Medium | Hard | Params(M) | GFLOPs |
---|---|---|---|---|---|---|
RetinaFace | MobileNet0.25 | 87.78 | 81.16 | 47.32 | 0.44 | 0.8 |
R50 | 94.92 | 91.90 | 64.17 | 30 | 38 | |
SCRFD | 0.5GF | 90.57 | 88.12 | 68.51 | 0.57 | 0.5 |
2.5GF | 93.78 | 92.16 | 77.87 | 0.67 | 2.5 | |
10GF | 95.16 | 93.87 | 83.05 | 4 | 10 | |
34GF | 96.06 | 94.92 | 85.29 | 10 | 34 | |
YOLO5Face | YOLOv5n-0.5 (ShuffleNetv2) | 90.76 | 88.12 | 73.82 | 0.45 | 0.6 |
YOLOv5n (ShuffleNetv2) | 93.61 | 91.52 | 80.53 | 1.73 | 2.1 | |
YOLOv5s | 94.33 | 92.61 | 83.15 | 7 | 6 | |
YOLOv5m | 95.30 | 93.76 | 85.28 | 21 | 18 |
- 32,303 images
- 393,703 faces
- Contains large variations in scale, pose, expression, occlusion and illumination
- 40% training 10% validation, 50% testing
- 61 scene categories