Light-weight face detection on Android with pytorch model
This is Android demo for LightWeightFaceDetector
We train a ultra light weight face detection and landmark align model which is only 643.7 k.
| Models | Easy | Medium | Hard | Model Size |
|---|---|---|---|---|
| mobilenetv1_0.25 | 0.91718 | 0.79766 | 0.3592 | 1.6 M |
| mobilenetv2_0.1 | 0.85330 | 0.68946 | 0.2993 | 1.3 M |
| mobilenetv3_small | 0.93419 | 0.83259 | 0.3850 | 2.8 M |
| efficientnet-b0 | 0.93167 | 0.81466 | 0.37020 | 16.4 M |
| mobilenetv2_0.1_lite_3fpn | 0.80379 | 0.716958 | 0.34924 | 426 k |
| mobilenetv2_0.1_lite_3pan | 0.83309 | 0.77181 | 0.5301 | 643.7 k |
| mobilenetv2_0.1_lite_4pan | 0.83965 | 0.791419 | 0.5964 | 813 k |
This demo is follow the Android resnet18 demo.
After pytorch 1.3, pytorch model can be easily deploy to android, see details for how to deploy ur pytorch model.
