reconstruct error estimates

1404 views
Skip to first unread message

Daniel Ferguson

unread,
Feb 23, 2012, 12:18:17 AM2/23/12
to ReconstructMe
Kinect realtime scanning worked (a bit slowly) on my intel i7 machine.
I have a better video card on order to improve my 3D scanning results,
and expect to get better results soon.

I scanned a sample part for comparison with our high resolution
structured light scanner, and found the error to be around 6mm max
over a 500mm object using the realtime scanning.

With the --record function and --replay --highres the error was
reduced to 4mm max. Actually on average, the surfaces were quite
close, I just found a few places where the surfaces were out of
alignment and increased the error.

For scanning people, I must admit, it works absolutely fantastic. The
surface tracking seemed to work the best on people, clothing, and
other complex shapes with lots of curvature and form.

On the other hand, a mechanical part mounted on a large rotating flat
table was almost impossible to capture well; the tracking was thrown
off by the high proportion of 'flat' area, and had trouble
reconstructing the object mounted on top. Maybe there is a filter for
removing this type of feature?

I'll post pics as soon as I figure out how.

Christoph Heindl

unread,
Feb 23, 2012, 1:13:12 AM2/23/12
to recons...@googlegroups.com

Am Donnerstag, 23. Februar 2012 06:18:17 UTC+1 schrieb Daniel Ferguson:
Kinect realtime scanning worked (a bit slowly) on my intel i7 machine.
I have a better video card on order to improve my 3D scanning results,
and expect to get better results soon.

Yes, you shoudl
 

I scanned a sample part for comparison with our high resolution
structured light scanner, and found the error to be around 6mm max
over a 500mm object using the realtime scanning. 

With the --record function and --replay --highres the error was
reduced to 4mm max. Actually on average, the surfaces were quite
close, I just found a few places where the surfaces were out of
alignment and increased the error.

We get similar figures here. The record-replay pattern has the advantage
that _every_ frame is processed, whereas the --realtime command can only
process as many frames as can be processed per second. So on a lower-
end graphics card you will loose a certain amount of frames per second.
Hence the differences in detail.
 

For scanning people, I must admit, it works absolutely fantastic. The
surface tracking seemed to work the best on people, clothing, and
other complex shapes with lots of curvature and form.

Great!
 

On the other hand, a mechanical part mounted on a large rotating flat
table was almost impossible to capture well; the tracking was thrown
off by the high proportion of 'flat' area, and had trouble
reconstructing the object mounted on top. Maybe there is a filter for
removing this type of feature?

Removing which feature? What helps with flat parts is to add something
complex in the background (clothing etc) and then cut-off the background.
In our queue of videos to make there is one, that will use the segmentation
shown here


to automatically remove complex backgrounds.
 
I'll post pics as soon as I figure out how.

You could use flickr or image shark, etc.

Best Christoph 

Daniel

unread,
Feb 25, 2012, 10:15:41 PM2/25/12
to ReconstructMe
Here are some images of comparison scans. I have since upgraded the
video card, to the nVidia GTX560 and it works fantastic. The faster
tracking improves the error and made it much friendlier to use:

This is the part I was using as a sample, as scanned by my structured
light scanner. I used this as a baseline as my error is around +/-
0.1mm.
https://lh3.googleusercontent.com/-uuTRkQ3FraI/T0mgasKpqjI/AAAAAAAAARc/2H29Taj_kFY/s641/SL_SCan.PNG

My first low resolution scan
https://lh6.googleusercontent.com/-pLA9I398z9g/T0mgZEjpkiI/AAAAAAAAARM/OMaIwOdAfXY/s680/Kinect_Scan2.PNG
Compared to the baseline:
https://lh3.googleusercontent.com/-4gY7Tb-SSns/T0mgYsWQNdI/AAAAAAAAARA/ezmWey368aQ/s733/Kinect_Error2.PNG

Another attempt in highres mode
https://lh5.googleusercontent.com/-FFI5CMkMs7U/T0mgafKN5mI/AAAAAAAAARU/uVQ3LfVPHg8/s671/Kinect_Scan_Highres.PNG
Compared to the baseline:
https://lh5.googleusercontent.com/-9Wddvcl_AFk/T0mgYdJqLUI/AAAAAAAAAQ8/gGo6vWnlqtM/s726/Kinect_Error_Highres.PNG

-Dan


Christoph Heindl

unread,
Feb 26, 2012, 2:04:18 AM2/26/12
to recons...@googlegroups.com
Hey Dan, thanks for sharing! Can we reuse those pictures (license?) for our homepage?

Christoph Heindl

unread,
Feb 26, 2012, 2:16:00 AM2/26/12
to recons...@googlegroups.com
What's the size of the object?

c.a...@gmail.com

unread,
Jul 8, 2014, 5:42:20 AM7/8/14
to recons...@googlegroups.com

Hi Christoph, Dan,

Very interesting.

Is there an accuracy improvement to expect with the Kinect v2.0 ?

Charles

Bill Culverhouse

unread,
Aug 3, 2014, 9:07:55 AM8/3/14
to recons...@googlegroups.com
I'm not sure that much has been said about the comparison between the TOF sensor
of the Kinect 2 and the Structured Light sensor of the K1 but the 2 does have a much
higher res color camera. And the back-end software for recognition is better.

But the scans produced from the K2 look about as good as the kinect 1. The primesense 1.09 
appears to do a slightly better job. At least visually, looking at scans from both. But that may
be that the software hasn't been tweaked to use a near-mode for the K2.

b



--
You received this message because you are subscribed to the Google Groups "ReconstructMe" group.
To unsubscribe from this group and stop receiving emails from it, send an email to reconstructm...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Reply all
Reply to author
Forward
0 new messages