Head pose estimation

[hemanthkorrapati]

Hi, I see that the head pose estimation is encoded in clm_model.params_global and clm_model.params_local which are being updated in CLMTracker::PDM::CalcParams and CLMTracker::CLM::NU_RLMS. Is there a way by which I can just retrieve the head pose estimation using the ground truth labelling of the landmark positions ? For example, like the algorithm 1 in your thesis which takes as input the parameters learned offline on the training data and the landmarks and outputs the pose information p.

Thank you

[TadasBaltrusaitis]

Hi,

You can do that using the PDM::CalcParams(Vec6d& out_params_global, const Mat_<double>& out_params_local, const Mat_<double>& landmark_locations, const Vec3d rotation) function. You can pass it the ground truth location in the landmark_locations matrix (and optional initial rotation estimate), and it will return out_params_global which will include scaling, translation, rotation parameters, which can be used with camera parameters to get a head pose estimate (or used directly, but that will not account for perspective transform).

Thanks, Tadas

[hemanthkorrapati]

Hi, Thanks for the quick answer. I will try this out. Just to confirm, your approach does not account the perspective transform in visualizing the tracking results, right ? In other words, your landmark detection and tracking process does not encode or somehow calculate the perspective information. Am I correct ?

Thank you

[TadasBaltrusaitis]

The landmark detection does not account for the perspective transform explicitly, neither does the CalcParams, but the Vec6d pose_estimate_to_draw = CLMTracker::GetCorrectedPoseWorld(clm_model, fx, fy, cx, cy); does as it uses the camera focal lengths and optical centers.

-Tadas

[hemanthkorrapati]

Ah, I get the complete picture now :+1: Thanks so much :)

[hemanthkorrapati]

Hi Tadas,

I tried to use the CalcParams() function in a standalone fashion. But I am not getting the correct head pose. I am sure that I am doing something wrong. Could you take a look at the minimal code below and give your opinion ?

Thanks.

#include "CLM_core.h"

#include <fstream>
#include <sstream>

#include <boost/accumulators/accumulators.hpp>

using namespace std;
using namespace cv;

int main (int argc, char **argv)
{

    // Assuming 640x490 images
    float fx = 500, fy = 500, cx = 320, cy = 245;

        //Image sequence read from a file
    std::ifstream imageFile("Path_Images.txt");
    if(!imageFile.good())
    {
        std::cout << "Image list not found." << std::endl;
        return 0;
    }

    CLMTracker::PDM pdm;
    pdm.Read("./bin/model/pdms/In-the-wild_aligned_PDM_68.txt");

    string line;
    cv::Vec6d pose_params = 0;
    while(getline(imageFile, line))
    {
        Mat im = cv::imread(line, CV_LOAD_IMAGE_COLOR);
        line.replace(line.size()-3, 3, "pts");

        cv::Mat_<double> landmarks(136, 1);
        std::ifstream landmarkFile(line.c_str());
        if(!landmarkFile.good())
        {
            std::cout << "Landmark file: " << line << " not readable" << std::endl;
            return 0;
        }
        std::cout << "Processing Image: " << line << std::endl;
        string temp;
        landmarkFile >> temp;
        landmarkFile >> temp;
        landmarkFile >> temp;
        for(int i=0; i< 68; i++)
        {
            landmarkFile >> landmarks.at<double>(i, 0);
            landmarkFile >> landmarks.at<double>(i+68, 0);
        }

        cv::Mat_<double> local_params(34, 1);
        local_params.setTo(0);
        cv::Vec3d rotation;
        rotation[0] = pose_params[3]; rotation[1] = pose_params[4]; rotation[2] = pose_params[5];
        auto t1 = std::chrono::high_resolution_clock::now();
        pdm.CalcParams( pose_params, local_params, landmarks, rotation);
        auto t2 = std::chrono::high_resolution_clock::now();
        auto timeTaken = std::chrono::duration_cast<std::chrono::microseconds>(t2 - t1);
        std::cout << pose_params << std::endl;

        int thickness = 2;
        CLMTracker::DrawBox(im, pose_params, Scalar(0, 0, 255), thickness, fx, fy, cx, cy);
        imshow("tracking_result", im);
        waitKey(1);
    }

    return 0;
}

[hemanthkorrapati]

Never mind. Figured it out and works like a charm :) Thanks

[vijaychav]

Hi Hemanth, I know that its been a while, but could you point out what the error was. I was able to extract the 68 feature points but need some help on getting the pose parameters from those points.

Thanks!

[misakibiu]

Hi,Tadas, I am confused about the world coordinate, what is the original point of this world coordinate. And what about the head coordinate space? I could not find any information about it. Thanks!^O^

Jane