Ncnn: mtcnn人脸检测模型更新ncnn版本之后，pnet的forward结果与之前的pnet的forward结果不一样

Created on 9 Feb 2018  ·  8Comments  ·  Source: Tencent/ncnn

问题描述

同一个工程之前使用的ncnn版本记得不太清楚了，貌似是下面两个版本中的一个
20170919和20171017，当时mat.data的类型还是float * ；
前两天更新了ncnn的版本，自己编译的libncnn.a文件；发现之前的模型用了之后检测结果不对，做了几个实验，用了 caffe的模型 转为ncnn模型:

1.正确的结果
操作方式：旧版的ncnn，使用相对版本的caffe2ncnn直接转换caffe模型，得出正确的结果

2.错误结果a
操作方式：最新代码编译出来ncnn库，以及相对应的caffe2ncnn工具转换直接转换"caffe的模型"，最后跑出来的结果与 1 中结果不一致；

3.错误结果b
操作方式：最新代码编译出来ncnn库，使用caffe的工具升级caffer模型，即下面命令升级caffe

upgrade_net_proto_text [old prototxt] [new prototxt]
upgrade_net_proto_binary [old caffemodel] [new caffemodel]

然后用相对应的caffe2ncnn工具转换直接转换 "升级后的caffe的模型"，最后跑出来的结果与 1 中结果不一致；而且是pnet的forward结果不一样；

现象：
3中pnet的forward结果的shape与 1中pnet的forward结果的shape是一样的；
但是当我遍历打印mat的值时，发现其中的值都是不一样的

打印的代码：

vector<FaceDetector::BBox> FaceDetector::Detect(ncnn::Mat &img_, const IMAGE_DIRECTION orient,
                                                int min_size, float scale_factor) {
    vector<BBox> finalBbox_;
    img = img_;
    img_w = img.w;
    img_h = img.h;
    img.substract_mean_normalize(mean_vals, norm_vals);

    //当时这里min_size是考虑到最小人脸多小不去做识别，优化速度来考虑的，给的40
    float minl = img_w < img_h ? img_w : img_h;
    int MIN_DET_SIZE = 12;
    float m = (float) MIN_DET_SIZE / min_size;
    minl *= m;
    float factor = scale_factor;
    int factor_count = 0;
    vector<float> scales_;
    while (minl > MIN_DET_SIZE) {
        if (factor_count > 0)m = m * factor;
        scales_.push_back(m);
        minl *= factor;
        factor_count++;
    }

    OrderScore order;
    int count = 0;

    for (size_t i = 0; i < scales_.size(); i++) {
        int hs = (int) ceil(img_h * scales_[i]);
        int ws = (int) ceil(img_w * scales_[i]);
        ncnn::Mat in;
        resize_bilinear(img_, in, ws, hs);
        ncnn::Extractor ex = pnet.create_extractor();
        ex.set_light_mode(true);
        ex.set_num_threads(OMP_NUM);
        ex.input("data", in);
        ncnn::Mat score_, location_;
        ex.extract("prob1", score_);
        ex.extract("conv4-2", location_);
        LOGI("first score_.shape=( %d, %d, %d) location_.shape=( %d, %d, %d)\n",score_.h,score_.w, score_.c, location_.h,location_.w, location_.c);
        std::vector<BBox> boundingBox_;
        std::vector<OrderScore> bboxScore_;
        generateBBox(score_, location_, boundingBox_, bboxScore_, scales_[i]);
//        LOGI("first %d boundingBox_.size=%d\n",i, boundingBox_.size());
        nms_cpu(boundingBox_, bboxScore_, nms_threshold[0], UNION);

        for (vector<BBox>::iterator it = boundingBox_.begin(); it != boundingBox_.end(); it++) {
            if ((*it).exist) {
                firstBbox_.push_back(*it);
                order.score = (*it).score;
                order.oriOrder = count;
                firstOrderScore_.push_back(order);
                count++;
            }
        }
        bboxScore_.clear();
        boundingBox_.clear();
    }
    //the first stage's nms
    if (count < 1)return finalBbox_;

    nms_cpu(firstBbox_, firstOrderScore_, nms_threshold[1], UNION);
    refineAndSquareBbox(firstBbox_, img_h, img_w);
    LOGE("firstBbox_.size()=%d\n", firstBbox_.size());
    ....
    //second statge
    ......

}

void FaceDetector::generateBBox(ncnn::Mat score, ncnn::Mat location, vector<BBox> &boundingBox_,
                                vector<OrderScore> &bboxScore_, float scale) {
    int stride = 2;
    int cellsize = 12;
    int count = 0;
    //score p
    float *p = score.channel(1);
    float *q = score.channel(0);
//    float *plocal = location.data;

    float *l0 = location.channel(0);
    float *l1 = location.channel(0);
    float *l2 = location.channel(0);
    float *l3 = location.channel(0);
    LOGI("cxy ==============localtion");
    for(int i=0; i< location.h; i++) {
        for (int j = 0; j < location.w; j++) {
            LOGI("cxy *l0=%f *l1=%f *l2=%f *l3=%f",*l0, *l1, *l2, *l3);
            l0++; l1++; l2++; l3++;
        }
    }
    LOGI("cxy ==============score");

    BBox bbox;
    OrderScore order;
    for (int row = 0; row < score.h; row++) {
        for (int col = 0; col < score.w; col++) {
            LOGI("cxy *p=%f *q=%f",*p, *q);
            if (*p > threshold[0]) {
                bbox.score = *p;
                order.score = *p;
                order.oriOrder = count;
                bbox.x1 = round((stride * col + 1) / scale);
                bbox.y1 = round((stride * row + 1) / scale);
                bbox.x2 = round((stride * col + 1 + cellsize) / scale);
                bbox.y2 = round((stride * row + 1 + cellsize) / scale);
                bbox.exist = true;
                bbox.area = (bbox.x2 - bbox.x1) * (bbox.y2 - bbox.y1);
                for (int channel = 0; channel < 4; channel++)
                    bbox.regreCoord[channel] = ((float *) location.channel(channel))[0];
                boundingBox_.push_back(bbox);
                bboxScore_.push_back(order);
                count++;
            }
            p++;
            q++;
//            plocal++;
        }
    }
}

步骤1打印的结果：

02-09 16:40:27.149 7237-7725/com.wx.album.demo I/FaceDetectActivity: 5 (r,g,b,a)=(-113,87,-122,-1)
02-09 16:40:27.149 7237-7725/com.wx.album.demo I/FaceDetectActivity: ---5 (r,g,b,a)=(6e7776ff)
02-09 16:40:27.149 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI:  data length=393216
02-09 16:40:27.149 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: native Detect----width*height=256*384
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: first score_.shape=( 53, 34, 2) location_.shape=( 53, 34, 4)
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy ==============localtion
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.010091 *l1=-0.010091 *l2=-0.010091 *l3=-0.010091
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.034791 *l1=-0.034791 *l2=-0.034791 *l3=-0.034791
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.017588 *l1=-0.017588 *l2=-0.017588 *l3=-0.017588
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.013194 *l1=-0.013194 *l2=-0.013194 *l3=-0.013194
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.008050 *l1=-0.008050 *l2=-0.008050 *l3=-0.008050
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.018748 *l1=-0.018748 *l2=-0.018748 *l3=-0.018748
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.054014 *l1=-0.054014 *l2=-0.054014 *l3=-0.054014
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.099514 *l1=-0.099514 *l2=-0.099514 *l3=-0.099514
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.115605 *l1=-0.115605 *l2=-0.115605 *l3=-0.115605
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.056714 *l1=-0.056714 *l2=-0.056714 *l3=-0.056714

步骤3打印的结果：

02-09 16:53:27.621 14330-19685/com.wx.album.demo I/FaceDetectActivity: 5 (r,g,b,a)=(-113,87,-122,-1)
02-09 16:53:27.621 14330-19685/com.wx.album.demo I/FaceDetectActivity: ---5 (r,g,b,a)=(6e7776ff)
02-09 16:53:27.621 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI:  data length=393216
02-09 16:53:27.621 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI: native Detect----width*height=256*384
02-09 16:53:27.645 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI: first score_.shape=( 53, 34, 2) location_.shape=( 53, 34, 4)
02-09 16:53:27.645 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy ==============localtion
02-09 16:53:27.645 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.064085 *l1=0.064085 *l2=0.064085 *l3=0.064085
02-09 16:53:27.646 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.066753 *l1=0.066753 *l2=0.066753 *l3=0.066753
02-09 16:53:27.646 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.069597 *l1=0.069597 *l2=0.069597 *l3=0.069597
02-09 16:53:27.646 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.013377 *l1=0.013377 *l2=0.013377 *l3=0.013377
02-09 16:53:27.646 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.000714 *l1=-0.000714 *l2=-0.000714 *l3=-0.000714
02-09 16:53:27.646 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.019337 *l1=-0.019337 *l2=-0.019337 *l3=-0.019337
02-09 16:53:27.646 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.025768 *l1=-0.025768 *l2=-0.025768 *l3=-0.025768
02-09 16:53:27.646 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.047746 *l1=-0.047746 *l2=-0.047746 *l3=-0.047746
02-09 16:53:27.646 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.007314 *l1=-0.007314 *l2=-0.007314 *l3=-0.007314
02-09 16:53:27.646 14330-19685/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.062976 *l1=0.062976 *l2=0.062976 *l3=0.062976

还请楼主帮忙看看什么情况？ 谢谢

picture cxy200927099

Most helpful comment

今天又仔细检测了下代码，确实是自己没改完全，我后面的rnet和onet取结果方式没改过来，抱歉！
不过还是有点不明白
为什么
pnet 的输出依然使用 out.channel(i)
rnet 和 onet 输出 *(out.data + out.cstep * i) 改为 out[i]

picture cxy200927099 on 11 Feb 2018
👍3

All 8 comments

pnet 的输出依然使用 out.channel(i)
rnet 和 onet 输出 *(out.data + out.cstep * i) 改为 out[i]

github上原始MTCNN是matlab训练的 col-major 模型，所以结果不正确
https://github.com/Tencent/ncnn/wiki/FAQ-ncnn-produce-wrong-result

nihui picture nihui on 9 Feb 2018

嗯嗯，刚刚测试了下，确实是那个原因，我已经按照楼主说的做了测试，但是还是发现结果有点差异，最后pnet, rnet,onet的找到的bindingBox数量虽然与之前老版本中的数量一致；
但是发现找到的人脸不准，仔细查看了下forward的结果，发现有细微差别
原来正确的结果：

02-09 16:40:27.149 7237-7725/com.wx.album.demo I/FaceDetectActivity: 5 (r,g,b,a)=(-113,87,-122,-1)
02-09 16:40:27.149 7237-7725/com.wx.album.demo I/FaceDetectActivity: ---5 (r,g,b,a)=(6e7776ff)
02-09 16:40:27.149 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI:  data length=393216
02-09 16:40:27.149 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: native Detect----width*height=256*384
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: first score_.shape=( 53, 34, 2) location_.shape=( 53, 34, 4)
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy ==============localtion
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.010091 *l1=-0.010091 *l2=-0.010091 *l3=-0.010091
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.034791 *l1=-0.034791 *l2=-0.034791 *l3=-0.034791
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.017588 *l1=-0.017588 *l2=-0.017588 *l3=-0.017588
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.013194 *l1=-0.013194 *l2=-0.013194 *l3=-0.013194
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.008050 *l1=-0.008050 *l2=-0.008050 *l3=-0.008050
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.018748 *l1=-0.018748 *l2=-0.018748 *l3=-0.018748
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.054014 *l1=-0.054014 *l2=-0.054014 *l3=-0.054014
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.099514 *l1=-0.099514 *l2=-0.099514 *l3=-0.099514
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.115605 *l1=-0.115605 *l2=-0.115605 *l3=-0.115605
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.056714 *l1=-0.056714 *l2=-0.056714 *l3=-0.056714
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.039420 *l1=0.039420 *l2=0.039420 *l3=0.039420
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.165425 *l1=0.165425 *l2=0.165425 *l3=0.165425
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.146497 *l1=0.146497 *l2=0.146497 *l3=0.146497
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.062257 *l1=0.062257 *l2=0.062257 *l3=0.062257
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.105753 *l1=0.105753 *l2=0.105753 *l3=0.105753
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.091465 *l1=0.091465 *l2=0.091465 *l3=0.091465
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.105839 *l1=0.105839 *l2=0.105839 *l3=0.105839
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.087600 *l1=0.087600 *l2=0.087600 *l3=0.087600
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.123140 *l1=0.123140 *l2=0.123140 *l3=0.123140
02-09 16:40:27.174 7237-7725/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.058920 *l1=0.058920 *l2=0.058920 *l3=0.058920

找不到人脸的结果：

02-09 19:59:56.944 20946-21405/com.wx.album.demo I/FaceDetectActivity: 5 (r,g,b,a)=(-113,87,-122,-1)
02-09 19:59:56.944 20946-21405/com.wx.album.demo I/FaceDetectActivity: ---5 (r,g,b,a)=(6e7776ff)
02-09 19:59:56.944 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI:  data length=393216
02-09 19:59:56.944 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: native Detect----width*height=256*384
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: first score_.shape=( 53, 34, 2) location_.shape=( 53, 34, 4)
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy ==============localtion
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.010091 *l1=-0.010091 *l2=-0.010091 *l3=-0.010091
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.034791 *l1=-0.034791 *l2=-0.034791 *l3=-0.034791
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.017589 *l1=-0.017589 *l2=-0.017589 *l3=-0.017589
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.013194 *l1=-0.013194 *l2=-0.013194 *l3=-0.013194
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.008050 *l1=-0.008050 *l2=-0.008050 *l3=-0.008050
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.018748 *l1=-0.018748 *l2=-0.018748 *l3=-0.018748
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.054014 *l1=-0.054014 *l2=-0.054014 *l3=-0.054014
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.099514 *l1=-0.099514 *l2=-0.099514 *l3=-0.099514
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.115604 *l1=-0.115604 *l2=-0.115604 *l3=-0.115604
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=-0.056714 *l1=-0.056714 *l2=-0.056714 *l3=-0.056714
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.039420 *l1=0.039420 *l2=0.039420 *l3=0.039420
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.165425 *l1=0.165425 *l2=0.165425 *l3=0.165425
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.146497 *l1=0.146497 *l2=0.146497 *l3=0.146497
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.062257 *l1=0.062257 *l2=0.062257 *l3=0.062257
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.105753 *l1=0.105753 *l2=0.105753 *l3=0.105753
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.091465 *l1=0.091465 *l2=0.091465 *l3=0.091465
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.105839 *l1=0.105839 *l2=0.105839 *l3=0.105839
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.087600 *l1=0.087600 *l2=0.087600 *l3=0.087600
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.123140 *l1=0.123140 *l2=0.123140 *l3=0.123140
02-09 19:59:56.966 20946-21405/com.wx.album.demo I/WX_FACE_DETECT_JNI: cxy *l0=0.058920 *l1=0.058920 *l2=0.058920 *l3=0.058920

不知道是不是这个原因导致最后的输出结果不准确

cxy200927099 picture cxy200927099 on 9 Feb 2018

caffe2ncnn_mtcnn.cpp.txt
已经按照楼主所说方式改了下面pnet，rnet，onet的取值方式

pnet 的输出依然使用 out.channel(i)
rnet 和 onet 输出 *(out.data + out.cstep * i) 改为 out[i]

另外我用的模型是 ElegantGod作者提供的 ；之前正确的模型貌似也是这个，而且也参照了工程里面的代码caffe2ncnn_mtcnn.cpp mergy到最新的caffe2ncnn.cpp中，修改后的文件参照附件caffe2ncnn_mtcnn.cpp.txt

cxy200927099 picture cxy200927099 on 9 Feb 2018

今天又仔细检测了下代码，确实是自己没改完全，我后面的rnet和onet取结果方式没改过来，抱歉！
不过还是有点不明白
为什么
pnet 的输出依然使用 out.channel(i)
rnet 和 onet 输出 *(out.data + out.cstep * i) 改为 out[i]

cxy200927099 picture cxy200927099 on 11 Feb 2018
👍3

@cxy200927099 你好 ，我下载了最新的ncnn，使用了你提供的caffe2ncnn_mtcnn.cpp.txt转换工具
代码修改了rnet onet如下
if(score.channel(1)[0] > threshold[2]) 修改为 if(score[0] > threshold[2])
it->regreCoord[channel]=bbox.channel(channel)[0]; 修改为 it->regreCoord[channel]=bbox[channel];
结果人脸框个数和之前ElegantGod作者提供的是一样的，但是位置不对。是我代码还没改完吗？

Oliverwang picture Oliverwang on 27 Feb 2018

mtcnn代码更新新的ncnn后，在ncnn::Mat类型的赋值， img = img_ 出现错误是什么原因？

lvpchen picture lvpchen on 15 Mar 2018

你好，想请问下你pnet的预测结果是怎么解析的

WZJAI2018 picture WZJAI2018 on 11 Nov 2018

https://github.com/moli232777144/mtcnn_ncnn
可参考这个项目

nihui picture nihui on 18 Dec 2018
Was this page helpful?
0 / 5 - 0 ratings

Related issues

yizhaoyanbo picture yizhaoyanbo  ·  3Comments
ghost picture ghost  ·  3Comments
kame-lqm picture kame-lqm  ·  4Comments
lovehuanhuan picture lovehuanhuan  ·  3Comments
AlexeyAB picture AlexeyAB  ·  3Comments