zuoqing1988 / zqcnn-mtcnn-vs-libfacedetection Goto Github PK

非常感谢您开源的这套实现，我之前跑过一些MTCNN的性能数据，是基于NCNN版本的MTCNN，运行在ARM cortex-A72核上。在我这里，Cortex-A72单线程，对1920x1080的图片做人脸检测的结果是126ms. 我的A72 core的主频是1.8GHz。不知道您有没有和NCNN版本的MTCNN做过对比？

在SampleMTCNN.cpp程序中修改“mtcnn.SetPara(image0.cols, image0.rows, 20, 0.5, 0.6, 0.8, 0.4, 0.5, 0.5, 0.709, 3, 20, 4, special_handle_very_big_face)”第三个参数为80，会显著增加识别是速度，ZQ_CNN_MTCNN_old.h的类中“void SetPara(int w, int h, int min_face_size = 60, float pthresh = 0.6, float rthresh = 0.7, float othresh = 0.7,
float nms_pthresh = 0.6, float nms_rthresh = 0.7, float nms_othresh = 0.7, float scale_factor = 0.709,
int pnet_overlap_thresh_count = 4, int pnet_size = 12, int pnet_stride = 2, bool special_handle_very_big_face = false)”的这个函数中的第三个参数是怎么影响识别速度的？

您好，我在测试您的SampleMTCNN的时候，出现D:\ZQCNN-MTCNN-vs-libfacedetection-master\ZQCNN\x64\Debug>SampleMTCNN.exe 4_128x96.jpg 10 1 failed to open file model/det1-dw20-fast.zqparams的错误，我仔细检查了模型的名称，都没有问题，请问该如何解决？谢谢！

你好，
看了你的对比，确实感觉速度比较快，但是你对比的两个模型，都没有给出在相应测试集上的准确率和召回率，这个有进行过对比吗？谢谢

想要修改SampleMTCNN程序为1秒识别多少张680x480的图片，作者以前做过这的程序修改吗，没有做过可以给出一下修改思路和建议吗？

您好，我在PC和arm-cortex-A9 (32bit) 上分别测试了一下这个模型。PC上的效果很赞，但是不知道为何arm上的效果不佳，我怀疑是编译问题还是在arm上需要对网络进行特殊配置？

我使用的arm交叉编译命令为：

$ cmake .. -DCMAKE_BUILD_TYPE=RELEASE -DCMAKE_TOOLCHAIN_FILE=../Toolchain-arm-linux-gnueabihf.cmake -DSIMD_ARCH_TYPE=arm

在arm上运行的结果如下（输入图像为images中的4.jpg，没到找到face）：

$ ./SampleMTCNN 4.jpg 
rnet = 0.5 M, onet = 2.1 M
convert cost: 32.968 ms
Pnet [0]: resolution [640x480], resize:0.002 ms, cost:433.564 ms
Pnet [1]: resolution [454x341], resize:45.522 ms, cost:219.581 ms
Pnet [2]: resolution [322x242], resize:23.582 ms, cost:106.983 ms
Pnet [3]: resolution [229x172], resize:14.352 ms, cost:52.163 ms
Pnet [4]: resolution [162x122], resize:8.938 ms, cost:24.974 ms
Pnet [5]: resolution [115x86], resize:4.032 ms, cost:11.189 ms
Pnet [6]: resolution [82x61], resize:3.582 ms, cost:5.784 ms
Pnet [7]: resolution [58x44], resize:1.166 ms, cost:2.544 ms
Pnet [8]: resolution [41x31], resize:0.615 ms, cost:1.205 ms
Pnet [9]: resolution [29x22], resize:0.294 ms, cost:0.560 ms
Pnet [10]: resolution [27x20], resize:0.239 ms, cost:0.457 ms
nms cost: 0.494 ms, (20-->2)
nms cost: 0.254 ms, (17-->0)
nms cost: 0.197 ms, (35-->5)
nms cost: 0.190 ms, (45-->5)
nms cost: 0.123 ms, (37-->4)
nms cost: 0.059 ms, (16-->4)
nms cost: 0.015 ms, (2-->1)
nms cost: 0.004 ms, (0-->0)
nms cost: 0.003 ms, (0-->0)
nms cost: 0.002 ms, (0-->0)
nms cost: 0.011 ms, (2-->1)
nms cost: 0.060 ms
first stage candidate count: 22
stage 1: cost 965.940 ms
run Rnet [12] times, candidate after nms: 0 
stage 2: cost 45.866 ms
run Onet [0] times, candidate before nms: 0 
stage 3: cost 0.013 ms
final found num: 0
total cost: 1044.997 ms (P: 998.948 ms, R: 45.991 ms, O: 0.058 ms)
total 1.045 s / 1 = 1045.050 ms
num face: 0

左博士，您好，我想尝试一下使用mxnet对您提供的模型进行载入，mxnet版本为1.2.0cpu版本，结果报了个错误：
Traceback (most recent call last):
File "compile.py", line 10, in
sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
File "/home/mirror/.conda/envs/tf/lib/python3.5/site-packages/mxnet/model.py", line 419, in load_checkpoint
symbol = sym.load('%s-symbol.json' % prefix)
File "/home/mirror/.conda/envs/tf/lib/python3.5/site-packages/mxnet/symbol/symbol.py", line 2535, in load
check_call(_LIB.MXSymbolCreateFromFile(c_str(fname), ctypes.byref(handle)))
File "/home/mirror/.conda/envs/tf/lib/python3.5/site-packages/mxnet/base.py", line 149, in check_call
raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: [03:34:34] src/operator/custom/custom.cc:104: Check failed: CustomOperator::Get()->Find(params.op_type) != nullptr Cannot find custom operator negativemining_onlylandmark106

载入代码如下：
import mxnet as mx
from mxnet import ndarray as nd

prefix, epoch = "lnet106-big", 0
sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
image_size = (48, 48)
opt_level = 3

shape_dict = {'data':(1, 3, *image_size)}
nnvm_sym, nnvm_params = nnvm.frontend.from_mxnet(sym, arg_params, aux_params)

您能提供一下帮助吗？谢谢左博啦~

windows下mklml.dll如何生成？或者指出参考教程。

谢谢