3D reconstruction of a table scene using rtab-mapping

[Tracy]

Hi all

Recently, I want to make a 3D reconstruction for a table scene like many cans, bottles and cylinder in a table only using one RGBD camera. Is it possible to get some very dense and accurate reconstruction results of table scene? How could I set the parameters of RTAB mapping? Any ideas?

In addition, if I use Kinect2 instead of Xtion or Kinect one. Could I get some better reconstruction results?

Many thanks for your helps.

Best Wishes

[matlabbe]

Hi Tracy,

RTAB-Map is tuned for large-scale mapping, not for precise dense cloud of small objects. Could you give some examples of the table environment? In general, the kinect2 is more precise at longer ranges, but not sure for objects under 3 meters in comparison to Xtion or Kinect1.

If you are using the standalone version (Preferences -> 3D Rendering panel), you can create more points by decreasing decimation parameter to 1 instead of 4. However if you want to reduce GPU power required to show as many points, you can do it already in Export when regenerating point clouds option is enabled.

cheers

[Tracy]

Hi Matlabbe

Many thanks for your help every time.

Our table scene includes some bottles, cans, and cylinders. The length of object is about 0.3-0.5 meter.

We set parameters: decimation 1 and maximum distance 1m . The point cloud is performed by outlier filter, smoothing. Finally, we got some very good 3D object models.

 

We found that: Xtion can get more accurate 3D model for the object in about 0.5-1.5 meter. Kinect2 are more accurate for long distance reconstruction like reconstruction the whole room. When we use kinect2 for 3D reconstruction for the object in near place, there are always many outlier and missed point clouds in the object.

Do you have some suggestions?

Best Wishes  

[matlabbe]

Hi,

Can you show screenshots comparing Xtion vs Kinect v2 for the same object?

On a side note, if your objects have a lot of geometry (and not too many visual features), you could activate ICP to get a more precise registration between frames. To do so, you should activate "Generate laser scan from depth image" under "Laser scans" section in Source panel. In "Motion Estimation" panel (Advanced options), select "Visual + Geometry ICP" registration approach. The ICP parameters are under "ICP Registration" panel for fine tuning (e.g., you may want to activate Point-to-Plane ICP as you are mapping surfaces).

cheers,
Mathieu

[Tracy]

Hi Mathieu

Many thanks for your suggestions. We will try it.

Here is the 3D model from Xtion

https://www.dropbox.com/s/slefoq0bqpr6ycf/xtion_cloud_0.3_smooth_filter.ply?dl=0

Here is the 3D model from Kinect2. You can see there are some outliers and missing point there.

https://www.dropbox.com/s/cjsjfcpnnaet5fp/kiect2_cloud_0.3_smooth_filter.ply?dl=0

Many thanks

[matlabbe]

For the noisy points alone in the empty space, you can activate "Cloud Filtering" when exporting (e.g., increase the number of minimum neighbors to filter more points):