First prototype of computeCoordinates of point cloud, yet to be tested

[Tabjones]

I started the PR mentioned in #392 so you can see the changes, i'll test the function tomorrow and let you know. You can wait to merge until some tests are performed.
Cheers

[xlz]

I don't think you need to use C++11 features for this.

[Tabjones]

what you mean ? the use of STL array is overkill ?!

[xlz]

the use of STL array is overkill

Yes

[Tabjones]

Ok, you are probably right, i'll remove std::array.
Those pointers will leak yes and also the idx is wrong as you said, will probably cause segfault.
Didn't have time to test/fix those, but i noticed them.

Also i was thinking that probably it is more efficient to create a function that computes coordinates of a single point, given the indexs. something like this:
void computeCoordinate (Frame* undistorted, Frame* registered, int row, int col, float& x, float& y, float& z, float& rgb) So that the user can fill the point cloud with a single loop. Returning an array would force the user to reloop them all in his program to fill PointXYZRGB data.
Also more user-safe in terms of memory leaks, cause the user would not have to use float*.

If you agree i'll modify it accordingly

[xlz]

Starting with a basic one should be good.

void getPointXYZRGB(Frame* undistorted, Frame* registered, int row, int col, float& x, float& y, float& z, unsigned int& rgb)

[Tabjones]

Tried it and it looks acceptable.

Had to convert rgb into float, so that it can be directly assigned to pcl::PointXYZRGB::rgb.
There's lot of noise around borders and color-depth registration doesnt look precise, but i'm satisfied for now.

[floe]

Noise around borders is a known artifact of TOF cameras AFAIK. The registration might also get a bit better once #379 has been merged.

[moonsh]

Firstly thanks for your hard work.
I would like to access X,Y,Z values so I made code like this

for (int xi = 0; xi < 512; xi++)
{
    for (int yi = 0; yi < 424; yi++)
    {
registration->getPointXYZRGB(&undistorted, &registered, xi, yi, X, Y, Z, RGB);
SaveFile2 << "v   " << X << "   " << Y << "   " << Z << std::endl;
    }
}

After that I saved in Obj file and I opened it in cloudcompare. However the result is not good.
Is this right approach to access XYZ values ?

[Tabjones]

i think you inverted rows and columns: i mean what you called xi should iterate through rows, i.e. the height (424). Something like this:

  for (int r=0; r<424; ++r)
    for (int c=0; c<512; ++c)
      registration->getPointXYZRGB(undistorted, registered, r,c, x,y,z,rgb);

[moonsh]

Thanks . Here I have one more question. Image described by points is not dense and even rough. Is it maximum resolution kinect v2 can do ? or is there anyway to Increase resolution ?

[Tabjones]

Yes it is not dense, cause it could contain NaN values.
About resolution, I believe that is the sensor provided resolution for depth (512x424), because that is the actual resolution of the IR camera.
In principle you could increase resolution by performing upsampling and/or other smoothing filters of the depth image, but of course you'll have to reapply a modified registration of color-depth again.

If that is what you need there's another package (that uses libfreenect2) that does exactly this upsampling, giving you point clouds at quarter-hd or full hd (1920x1200). https://github.com/code-iai/iai_kinect2, definitely worth a try.
Cheers

[moonsh]

Ah I see ! Sorry for annoying you again . How about RGB values ? For visualization, R,G and B values are needed separately but here I can get only rgb value .

[Tabjones]

umh, not sure but you could try unpacking as described here http://docs.pointclouds.org/1.7.0/structpcl_1_1_point_x_y_z_r_g_b.html

[moonsh]

Here is result I got but some problem is exist . As you can see some points are missing in above figure. What factors makes this missing points? too near ? or Am I taking point data too quick ?

[xlz]

@moonsh Maybe reflection is too strong and exceeds the sensing limit. Some of the low quality data is also removed from output.

I have also seen flat surface with patches without data. I wonder if there is any way to get around this. Maybe you can try using a filter to lower IR intensity.