1.libfacedetection
它最大的亮點是能夠在所有支持 C/C++的平臺上編譯運行。作者將預訓練的 CNN 模型轉換爲靜態變量後儲存到了 C 文件裏,使得該項目不需要任何其他依賴項(當然 OpenCV 還是需要的),僅僅只需要一個 C++編譯器,就能在任何一個平臺甚至嵌入式系統上編譯並運行該項目。
尤其吸引人的是該項目使用 C++編寫且支持 AVX2,在 i7 的 CPU 上就能跑出喪心病狂的 1000FPS!下圖爲項目作者給出的檢測效果示例。
WIDER Face上的性能
項目開源地址:https://github.com/ShiqiYu/libfacedetection
其他資料:https://mp.weixin.qq.com/s?__biz=MzUzODkxNzQzMw==&mid=2247484161&idx=1&sn=9e42c21deef5338570fc3cfcdcaaa007&chksm=fad12c57cda6a54196e9169a3d2c52fc9ad7d0977739b8405afdafea3e396f2afac8cef85b4d&mpshare=1&scene=1&srcid=&sharer_sharetime=1585005615041&sharer_shareid=d0cc0137df4356a868fcd0ba3aceb4b6&key=f506612bfbba42cbc076acb4867863377c726e29627947a03467baa275dc37afc31bbc11cf85764accd00292cac95833010751fa139c01a3fd85346d21a8e8595a9ec553fc36cffc4a41070339586026&ascene=1&uin=MTAxOTEyNTg4MQ%3D%3D&devicetype=Windows+10&version=62080079&lang=zh_CN&exportkey=AaVavDy7GmfajUgke2I%2FJh4%3D&pass_ticket=bQvMKORzh3yjq1zpONwP1ldTQRTNRs1d64U1UCD%2BoUhU6Qu4gKyZkRvA2gSmgGAb
已在win10下基於VS2017成功運行
2.Retinaface
https://arxiv.org/pdf/1905.00641.pdf
代碼地址:
https://github.com/deepinsight/insightface/tree/master/RetinaFace
各種版本:
1).這個6ms,能檢測近距離人臉,無關鍵點 mxnet
代碼地址:https://github.com/jacke121/faster-mobile-retinaface
後來又更新了:
代碼地址:https://github.com/1996scarlet/faster-mobile-retinaface
2).RetinaFace in PyTorch
代碼地址:https://github.com/biubug6/Pytorch_Retinaface
3).68-Retinaface-Pytorch-version
代碼地址:https://github.com/jacke121/68-Retinaface-Pytorch-version
4).移動端快速人臉檢測模型是基於RetinaFace的優化去掉stride8以及stride32和stride16的landmark 在CPU位Intel® Pentium® CPU G2020 @ 2.90GHz(2900 MHz)的設備中人臉檢測可達到40ms/幀
代碼地址:https://github.com/qaz734913414/ncnn_faster-mobile-retinaface
5).基於RetinaFace的目標檢測方法,適用於人臉、缺陷、小目標、行人等,號稱比原版還好,resnet50,
代碼地址:https://github.com/bleakie/RetinaDetector
6).fddb數據集
代碼地址:https://drive.google.com/file/d/17t4WULUDgZgiSy5kpCax4aooyPaz3GQH/view
7).這個權重更小:出5個關鍵點,好 300*255 32 110ms
代碼地址:https://github.com/biubug6/Pytorch_Retinaface
8).這兩個是resnet50,可以檢測別的:
代碼地址:https://github.com/bleakie/RetinaDetection
代碼地址:https://github.com/deepinsight/insightface/tree/master/RetinaFace
9).這個13ms,但是誤檢率很高,無關鍵點。
代碼地址:https://github.com/jacke121/Ultra-Light-Fast-Generic-Face-Detector-1MB
10).cpu版,100ms,無關鍵點,嚴重側臉也能檢測:
代碼地址:https://github.com/Star-Clouds/centerface
**11) c++ _caffe實現:
MXNet轉caffe
代碼地址:https://github.com/clancylian/retinaface
完整版caffe-cpp實現
代碼地址:https://github.com/wzj5133329/retinaface_caffe(已實現)
12)這個權重113m,出3個角度方向:
代碼地址:https://github.com/supernotman/RetinaFace_Pytorch
13).人臉方向是個網絡,90多m權重 :
代碼地址:https://github.com/natanielruiz/deep-head-pose