Ask Your Question

dnn module - Face Detection - poor results - Open CV 3.4.3 [closed]

asked 2018-11-13 02:42:37 -0600

rc gravatar image

updated 2020-09-07 04:58:49 -0600

I am getting poor results with the DNN face detection module for images in which the Haarscard cascade works fine.

C++ Code is as follows:

#include <iostream>
#include <string>
#include <vector>
#include <stdlib.h>
#include <opencv2/core.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/dnn.hpp>

using namespace cv;
using namespace std;
using namespace cv::dnn;

const size_t inWidth = 300;
const size_t inHeight = 300;
const double inScaleFactor = 1.0;
const float confidenceThreshold = 0.7;
const cv::Scalar meanVal(104.0, 177.0, 123.0);

#define CAFFE

const std::string caffeConfigFile = "deploy.prototxt";
const std::string caffeWeightFile = "res10_300x300_ssd_iter_140000_fp16.caffemodel";

const std::string tensorflowConfigFile = "opencv_face_detector.pbtxt";
const std::string tensorflowWeightFile = "opencv_face_detector_uint8.pb";

void detectFaceOpenCVDNN(Net net, Mat &frameOpenCVDNN)
    int frameHeight = frameOpenCVDNN.rows;
    int frameWidth = frameOpenCVDNN.cols;

    //resize(frameOpenCVDNN, frameOpenCVDNN, Size(300, 300));

#ifdef CAFFE
        cv::Mat inputBlob = cv::dnn::blobFromImage(frameOpenCVDNN, inScaleFactor, cv::Size(inWidth, inHeight), meanVal, false, false);
        cv::Mat inputBlob = cv::dnn::blobFromImage(frameOpenCVDNN, inScaleFactor, cv::Size(inWidth, inHeight), meanVal, true, false);

    net.setInput(inputBlob, "data");
    cv::Mat detection = net.forward("detection_out");

    cv::Mat detectionMat(detection.size[2], detection.size[3], CV_32F, detection.ptr<float>());

    for(int i = 0; i < detectionMat.rows; i++)
        float confidence =<float>(i, 2);
        cout << confidence << endl;

        if(confidence > confidenceThreshold)

            int x1 = static_cast<int>(<float>(i, 3) * frameWidth);
            int y1 = static_cast<int>(<float>(i, 4) * frameHeight);
            int x2 = static_cast<int>(<float>(i, 5) * frameWidth);
            int y2 = static_cast<int>(<float>(i, 6) * frameHeight);

            cv::rectangle(frameOpenCVDNN, cv::Point(x1, y1), cv::Point(x2, y2), cv::Scalar(0, 255, 0),2, 4);


int main( int argc, const char** argv )
#ifdef CAFFE
  Net net = cv::dnn::readNetFromCaffe(caffeConfigFile, caffeWeightFile);
  Net net = cv::dnn::readNetFromTensorflow(tensorflowWeightFile, tensorflowConfigFile);

  /*VideoCapture source;
  if (argc == 1);
  Mat frame = imread("barry2.jpg");
  cout << "Channels: " + to_string(frame.channels()) << endl;
  //Mat resized;
  //resize(frame, resized, Size(300, 300));

  double tt_opencvDNN = 0;
  double fpsOpencvDNN = 0;
      //source >> frame;
      //double t = cv::getTickCount();
      detectFaceOpenCVDNN ( net,frame);
      //tt_opencvDNN = ((double)cv::getTickCount() - t)/cv::getTickFrequency();
      //fpsOpencvDNN = 1/tt_opencvDNN;
      //putText(frame, format("OpenCV DNN ; FPS = %.2f",fpsOpencvDNN), Point(10, 50), FONT_HERSHEY_SIMPLEX, 1.4, Scalar(0, 0, 255), 4);
      imshow( "OpenCV - DNN Face Detection",frame);
      int k = waitKey(0);

Sample image attached for which i am failing to get results, unless i lower confidence factors down to about 0.1 (10%).

Any ideas?



edit retag flag offensive reopen merge delete

Closed for the following reason the question is answered, right answer was accepted by dkurt
close date 2018-11-14 09:48:51.154207

1 answer

Sort by ยป oldest newest most voted

answered 2018-11-13 03:20:13 -0600

dkurt gravatar image

This is your image after resizing to 300x300:

image description

Object detection networks can work with different input sizes. In example,

1296x864 (origin sizes):

image description

648x432 (x2):

image description

And so on...

edit flag offensive delete link more


Many thanks for the quick response. I am 100% certain that in a earlier variant of the code above I tried CV::resize with a size parameter of Size(300,300) before passing it to the function.

I believe this is shown in the commented out line

//resize(frameOpenCVDNN, frameOpenCVDNN, Size(300, 300));

I also tried reiszing it to 300 x 300 before calling the code.

Can you confirm you are using the code I posted above, as I am baffled as to what you have done differently to me! :-)

I am using the models posted at Is it possible that we are using different TF or caffemodels. Are you able to share whixh models you are using?


rc gravatar imagerc ( 2018-11-13 03:35:16 -0600 )edit

@rc, The results for the same Caffe's model. blobFromImage does resize internally so there is no need to call it separately.

dkurt gravatar imagedkurt ( 2018-11-13 03:49:35 -0600 )edit



Can you confirm you are using the same code as above , the caffe model posted here:

I am still getting no results using the code above when passing in the original unresized image unless I reduce the confidence factor to 0.1 at which point I get lots of false positives.

For the record, I am using opencv 3.4.3 on windows 8.1, not on a GPU.

I don't have the caffe or TF framworks installed, only opencv.

I am really confused!



rc gravatar imagerc ( 2018-11-13 04:09:22 -0600 )edit

@rc, Did you try to replace cv::Size(inWidth, inHeight) which is cv::Size(300, 300) to something else?

dkurt gravatar imagedkurt ( 2018-11-13 06:47:19 -0600 )edit

Dkurt, I believe so, but I can't be certain. I am away from my development machine for the remainder of the afternoon but will try it later today and give an update. Thanks for your continued patience!

rc gravatar imagerc ( 2018-11-13 07:26:10 -0600 )edit

Apologies dkurt. The replacement of the line cv::Size(inWidth, inHeight) which is cv::Size(300, 300) with the actual original image sizes worked perfectly. I was obviously wrong when i thought i had tried that before. Many thanks for your assistance, and apologies for such a basic mistake.

rc gravatar imagerc ( 2018-11-13 18:39:09 -0600 )edit

@rc, Never mind. Nice to help you!

dkurt gravatar imagedkurt ( 2018-11-14 09:48:21 -0600 )edit

I am using almost identical code but without gui. I have tested a number of images on different resolutions. Sometimes a number of false positives is detected outside the frame! The higher the resolution (of the same image) the more false positives. I use this: cv::Mat inputBlob = cv::dnn::blobFromImage(frameOpenCVDNN, inScaleFactor, cv::Size(frameWidth, frameHeight), meanVal, false, false); in which frameWidth and frameHeight are the size of the frame.

arnov gravatar imagearnov ( 2019-02-12 12:32:33 -0600 )edit

@arnov did you solve the issue with the false positives?

sfo gravatar imagesfo ( 2019-05-09 09:09:03 -0600 )edit
sturkmen gravatar imagesturkmen ( 2020-09-07 05:02:57 -0600 )edit

Question Tools

1 follower


Asked: 2018-11-13 02:42:37 -0600

Seen: 4,933 times

Last updated: Nov 13 '18