Questions about function ExtractDynPoints()

[KaninchenM]

Hi, sorry to bother you. I learned Semantic SLAM in recent months and now I'am doing some experiment basing on your code. After reading your paper and code, I have some confusions especially for the function Geometry::ExtractDynPoints(const vector<ORB_SLAM2::Frame> &vRefFrames,const ORB_SLAM2::Frame &currentFrame) (in Dynaslam/src/Geometry.cc).

1. What's the meaning of variable vMPw

   cv::Mat vMPw = refFrame.mTcw.inv() * vMPRefFrame;

2. What is the purpose of the judgment sentence

   const float d = refFrame.mImDepth.at<float>(v,u);
           if (d > 0 && d < 6){
               matRefFrame.at<float>(k,0) = refFrame.mvKeysUn[j].pt.x;
               matRefFrame.at<float>(k,1) = refFrame.mvKeysUn[j].pt.y;
               matRefFrame.at<float>(k,2) = 1.;
               matInvDepthRefFrame.at<float>(k,0) = 1./d;
               matDepthRefFrame.at<float>(k,0) = d;
               vLabels.at<float>(k,0) = i;
               k++;
           }

   why the d should (d > 0 && d < 6)?

   for (int i(0); i < matProjDepth.cols; i++)
       {
           if (matProjDepth.at<float>(0,i) < 7)
           {
               __matProjDepth.at<float>(0,h) = matProjDepth.at<float>(0,i);
   
               _vMPCurrentFrame.at<float>(0,h) = vMPCurrentFrame.at<float>(0,i);
               _vMPCurrentFrame.at<float>(1,h) = vMPCurrentFrame.at<float>(1,i);
               _vMPCurrentFrame.at<float>(2,h) = vMPCurrentFrame.at<float>(2,i);
               _vMPCurrentFrame.at<float>(3,h) = vMPCurrentFrame.at<float>(3,i);
   
               _vAllMatRefFrame.at<float>(h,0) = vAllMatRefFrame.at<float>(i,0);
               _vAllMatRefFrame.at<float>(h,1) = vAllMatRefFrame.at<float>(i,1);
               _vAllMatRefFrame.at<float>(h,2) = vAllMatRefFrame.at<float>(i,2);
   
               _vLabels.at<float>(h,0) = vLabels.at<float>(i,0);
   
               __vAllDepthRefFrame.at<float>(h,0) = vAllDepthRefFrame.at<float>(i,0);
   
               h++;
           }
       }

   and why matProjDepth.at<float>(0,i) < 7

3. How to set the value of mDmax?

   cv::Mat u1((2*mDmax+1)*(2*mDmax+1),2,CV_32F);
       int m(0);
       for (int i(-mDmax); i <= mDmax; i++){
           for (int j(-mDmax); j <= mDmax; j++){
               u1.at<float>(m,0) = i;
               u1.at<float>(m,1) = j;
               m++;
           }
       }

   and what's the meaning of mDmax,such as in

   if (matDepthDifference.at<float>(i,0) > mDepthThreshold)
           {
               int xIni = (int)matCurrentFrame.at<float>(i,0) - mDmax;
               int yIni = (int)matCurrentFrame.at<float>(i,1) - mDmax;
               int xEnd = (int)matCurrentFrame.at<float>(i,0) + mDmax + 1;
               int yEnd = (int)matCurrentFrame.at<float>(i,1) + mDmax + 1;
              ......
          }

4. How to set the threshold τz(in code, it's mDepthThreshold)? In paper, it's 0.4m and in code it's 0.6m

5. What's the use of mVarThreshold? Why only if (_stddev*_stddev)< mVarThreshold , can the keypoint be dynPoint?

6. I've statistics the spent time of GeometricModelCorrection part. In different data sequences, time-consuming varies greatly. In fr3_walking_rpy, it needs about 0.2s/frame while in fr3_walking_xyz, it needs 0.45+s/frame. Could you please give me some suggestions about what factors affect the GeometricModelCorrection part's spending time?

[KaninchenM]

Hope to get your help. Thank you so much!

[KaninchenM]

Answer Q5: find the idea in paper