原
MTCNN算法提速应用(ARM测试结果评估)
置顶
2017年11月02日 10:48:05
samylee
阅读数:11584
<span class="tags-box artic-tag-box">
<span class="label">标签:</span>
<a data-track-click="{"mod":"popu_626","con":"mtcnn人脸检测"}" class="tag-link" href="http://so.csdn.net/so/search/s.do?q=mtcnn人脸检测&t=blog" target="_blank">mtcnn人脸检测 </a>
<span class="article_info_click">更多</span></span>
<div class="tags-box space">
<span class="label">个人分类:</span>
<a class="tag-link" href="https://blog.csdn.net/samylee/article/category/6921813" target="_blank">人脸检测 </a>
</div>
</div>
<div class="operating">
</div>
</div>
</div>
</div>
<article>
<div id="article_content" class="article_content clearfix csdn-tracking-statistics" data-pid="blog" data-mod="popu_307" data-dsm="post">
<div class="article-copyright">
版权声明:本文为博主原创文章,欢迎转载。 https://blog.csdn.net/samylee/article/details/78421960 </div>
<link rel="stylesheet" href="https://csdnimg.cn/release/phoenix/template/css/ck_htmledit_views-f76675cdea.css">
<div class="htmledit_views">
<p>经博主测试,mtcnn原三层网络如果用于工程测试,<span style="color:#ff0000;">误检情况严重</span>,在fddb上测试结果也是,经常将<span style="color:#ff0000;">手或者耳朵</span>检测为人脸,这个很头疼(<span style="color:#ff0000;">因为标注数据!</span>),所以重新训练显得尤为重要!</p>
博主的改进方法及如何重新训练的就不具体介绍了,主要思想就是用卷积取代池化,fddb测试离散ROC88!
注意:某些公开的非官方mtcnn训练方法有误!只可参考,不可深入!
PC端测试:(测试软件:vs2015,测试硬件:i7-4790-4core)
1920x1080视频,最小检测人脸为60,速度为22ms!
640x480视频,最小人脸为25,速度为17ms!
arm端测试:(硬件:香橙派,全志A64芯片,4核64位Cortex-A53,市场价格240元!)
测试:640x480视频,最小检测人脸80,速度30ms!
测试效果如下(这里对比了Shiqi YU的人脸检测):
算法 |
测试图像尺寸 |
测试最小人脸尺寸 |
算法耗时(ms) |
ShiqiYU-facedetect_frontal |
2064x1078 |
40 |
95 |
ShiqiYU-facedetect_frontal_surveillance |
2064x1078 |
40 |
125 |
ShiqiYU-facedetect_multiview |
2064x1078 |
40 |
215 |
ShiqiYU-facedetect_multiview_reinforce |
2064x1078 |
40 |
380 |
OURS |
2064x1078 |
40 |
83 |
ShiqiYU-facedetect_fronta
ShiqiYU-facedetect_frontal_surveillancel
ShiqiYU-facedetect_multiview
ShiqiYU-facedetect_multiview_reinforce
ours
任何问题请加唯一QQ2258205918(名称samylee)!