seetaface / SeetaFaceEngine

Other
4.59k stars 1.78k forks source link

arm cortex-a9 Identification优化结果 #88

Open sylijinlei opened 7 years ago

sylijinlei commented 7 years ago

本人移植到了tiny4412上,FaceIdentificaton没经过Neon优化,识别时间为17.8秒,经过Neon优化后,识别速度为1.35秒;另外FaceDetection和FaceAlignment没经过优化,时间为120ms,220ms;

sylijinlei commented 7 years ago

cpu为arm-cortex-a9, 1.4ghz, 四核

chibai commented 7 years ago

这个face identification太蛋疼了!!看来没法本地化了

ujoker commented 7 years ago

@sylijinlei 怎么优化呢?有没有可参考的例程?对ARM不太熟,谢谢啦

szad670401 commented 7 years ago

我提取特征 安卓上 exynos7420 500ms左右. 小米3的tegra4 1.2秒左右

sunjunlishi commented 7 years ago

诺基亚 RM-980,1GHZ 512内存, 特征提取5秒; 华为手机x2 麒麟930,4A53@2.0GHz + 4A53@1.5GHz ,特征提取1.5秒吧;neon 改为 openblas 实现矩阵乘,速度提取特征变为400多ms

red010182 commented 7 years ago

樹莓派3 model B,1.2GHZ,1GB內存,無超頻,neon detect: 85ms align: 128ms identify: 2.822s

圖片是160x120

red010182 commented 7 years ago

iPhone 6, neon detect: 50ms align: 17ms identify: 0.45s

圖片是240x240