适用于移动端的人脸识别模型,计算量于mobilefacenet相同,但megaface上提升了2%+。我训练的mobilefacenet在megaface VER上比原作者 高出了2%,因为训练的数据和方法不一样。
Methods | Flops (112x112) | LFW | CFP-FP | AgeDB | Megaface-Id | Megaface-Ver@1e-6 | 备 注 |
---|---|---|---|---|---|---|---|
MobileFaceNet440,R | 440M | 99.70+ | 96.70+ | 96.95+ | 92.85+ | 94.20+ | 未开源 |
ZW350 | 356M | 99.70+ | 96.82+ | 97.00+ | 93.90+ | 94.70+ | 未开源 |
ZW400 | 404M | 99.70+ | 96.95+ | 97.00+ | 94.46+ | 95.60+ | 未开源 |
MobileFaceNet600,R | 612M | 99.76+ | 97.60+ | 97.50+ | 95.14+ | 95.98+ | 已开源 |
ZW440 | 444M | 99.76+ | 97.30+ | 97.40+ | 95.25+ | 96.00+ | 已开源 |
zw440-ver
Methods | Openvino | opencv单线程 |
---|---|---|
MobileFaceNet600,R | 6ms | 141ms |
ZW440 | 7ms | 80ms |
经过测试,zw440并没有Mobilefacenet600M快.感谢moli的测试
模型包含mxnet ncnn caffe 三种格式 Baidu Drive 提取码:b0dm
https://github.com/deepinsight/insightface/tree/master/iccv19-challenge
https://github.com/deepinsight/insightface
https://github.com/happynear/FaceVerification
https://github.com/Tencent/ncnn
https://github.com/cypw/MXNet2Caffe
没有做速度方面考虑,后期跟进改善。