What does re_object_recorder do?

The re_object_recorder package aims at the creation of object detection models using the Kinect sensor. It allows the user to control the recording process and improve the results by manually removing incorrect clouds. By default, 2 types of models with the following names are created for each object:

  • kinectmodel.zip: Contains all the data collected during the scanning process. This model files can be used by re_kinect_object_detector.

  • 2dcammodel.zip: A more space efficient object model. These files can be used by re_vision.

If a RoboEarth connection is established, you can also upload the object model files to the database.

See the tutorial roboearth_stack/Tutorials/Record an object model on how to use it.

The object under investigation has to be placed on a marker table, that holds the following pattern marker_template.pdf. It is important to print it out the same size (marker edge length of 80.0 mm). These AR (Augmented Reality) markers are used to register the point cloud for the model from different views without any expensive computations. All packages are located in the object_scanning subdirectory under the roboearth_stack.

This package relies on ar_bounding_box to create a bounding box fixed to the marker pattern. The bounding box is used to filter the points belonging to the model from background points.

Common Problems and Solutions

General Hints

If re_object_recorder is unable to receive object recordings, please consider the following hints:

  • Make sure that the Kinect camera is not too close to the marker pattern. A distance of around 60 cm from the Kinect to the marker pattern should be enough.
  • Check that the black boxes on the marker template are squares with a side length of 80 mm. You may need to disable printing margins on your printer settings to achieve this.
  • Make sure that the marker template is lit uniformly.
  • Glossy and poorly textured objects will probably give poor detection results.
  • If none of the hints listed here is helpful to you, consider writing a mail to our mailing list: https://lists.roboearth.org/mailman/listinfo/roboearth-users_roboearth.org

'detected X markers in the image; at least 3 markers are necessary to estimate a coordinate system. please check your camera setup.'

ar_bounding_box tries to detect the 3 markers closest to the Kinect and create a coordinate system; this warning gets displayed whenever there are less than 3 markers detected.

  • Make sure the marker template is lit uniformly. E.g. turn on the artificial light and try to remove light sources that could interfere with the recording process.
  • Make sure that the camera has 3 markers completely in its field of view. You can check this with e.g. the rviz tool.
  • Check that the point cloud produced by the Kinect contains the marker template (using the rviz tool). If not, try moving the camera a bit away from the marker template. The Kinect can only provide reliable depth information from a distance of around 60cm from the object under consideration (see also openni_kinect/kinect_accuracy).

'distance X/X too large: XXXX'

This message indicates that the ar_bounding_box node was unable to calculate a coordinate frame from the detected markers.

In order to eliminate implausible scans, we check the distances between the estimated marker positions. If they differ from the distances given by the marker template by more than a small threshold, ar_bounding_box outputs the above warning. In this case, you should check the estimated marker positions with rivz.

There are multiple ways to address this problem:

  • Check that your camera setup allows the Kinect to always have at least three markers fully in its view. You can check this with e.g. the rviz tool.
  • Remove objects that look similar to markers from the Kinect's field of view.
  • Try to improve the lighting conditions (see above).
  • Make sure the marker template has the right size (each marker must have an edge length of exactly 80 mm), is assembled precisely (the RoboEarth logo should form a perfect circle) and is completely planar.

  • As a last resort you can increase the 'marker_dist_threshold' parameter in ar_bounding_box/launch/ar_kinect.launch'. Increasing this parameter will however imply that the maximum allowable error on the pattern_base coordinate frame increases, hereby allowing larger deviations between measured scans.

Bad alignment of point clouds

If the point clouds do not seem to be aligned at all, make sure that depth registration is enabled for the openni driver. See openni_launch#Quick_start on how to do that.

Subscribed Topics

Wiki: re_object_recorder (last edited 2013-02-19 13:05:23 by DanielDiMarco)