Registration

[giacomodabisias]

Hi,
I am trying to apply the new registration you did with the new parameters taken directly from the kinect2.

I created the Registration object and took the parameters from the device with the
getIrCameraParams()
getColorCameraParams()

functions.
Now I have to use the apply function of the class, but I am wondering what size the two images should have. Should I resize the rgb to the depth size or can I use the function without resizing the images?

Thanks

[xlz]

#190 has demos showing how to use the registration class. Code is here https://github.com/larshg/libfreenect2/tree/registration .

[floe]

Thanks @xlz - I will hopefully get around to integrating these examples into the master branch soon (a bit swamped with work at the moment :-).

[giacomodabisias]

I tried the registration, but I assume there must be something wrong in the parameters probably (?)
Here is the cloud created with the registration:

while the one made with my own calibration looks like this:

I could create a small video to see the difference if needed.

[floe]

Can you describe in a bit more detail how they differ? They look quite similar to me... but in general, a proper chessboard calibration will usually have higher quality than the factory calibration, which is IMHO optimized for speed.

[giacomodabisias]

I will prepare a short video, so you can see better. If you enlarge the two
images and look at the hand you will see a lot of noise

2015-05-06 14:09 GMT+02:00 Florian Echtler notifications@github.com:

Can you describe in a bit more detail how they differ? They look quite
similar to me... but in general, a proper chessboard calibration will
usually have higher quality than the factory calibration, which is IMHO
optimized for speed.

—
Reply to this email directly or view it on GitHub
#223 (comment)
.

G. Dabisias, PhD Student
PERCRO, (Laboratory of Perceptual Robotics)
Scuola Superiore Sant'Anna
via Luigi Alamanni 13D, San Giuliano Terme 56010 (PI), Italy
mob.: +39 3480839095

[floe]

PR #226 contains a minimal implementation of the registration code; could you give that a try and see if it still looks wrong? Note that this is not yet using a depth buffer, so some pixel duplication is inevitable.

[giacomodabisias]

Hi,
I just checked it and it doesn't look too bad. I will recheck everything
and let you know.

Thanks

2015-05-08 12:55 GMT+02:00 Florian Echtler notifications@github.com:

PR #226 #226 contains a
minimal implementation of the registration code; could you give that a try
and see if it still looks wrong? Note that this is not yet using a depth
buffer, so some pixel duplication is inevitable.

—
Reply to this email directly or view it on GitHub
#223 (comment)
.

G. Dabisias, PhD Student
PERCRO, (Laboratory of Perceptual Robotics)
Scuola Superiore Sant'Anna
via Luigi Alamanni 13D, San Giuliano Terme 56010 (PI), Italy
mob.: +39 3480839095

[floe]

I've merged PR #223 with the basic example code, I'll use this issue to track the remaining features to be implemented:

• add depth buffer to avoid color pixel duplication
• always use 4-channel image as output (optionally using the 4th channel for downsampled depth data)
• add helper functions for point clouds (i.e. xyzRGB output)?
• add standalone depth undistortion

[xlz]

Is undistorted depth image required for generating point clouds?

[JuantAldea]

I'd say so, just to keep straight lines as straight as possible.