sdavies's profile - activity

I noticed that the kinect sensor in turtlebot_gazebo (in the unstable release) produces only black and white images. The gazebo.urdf.xacro file looks like it is supposed to produce color. Also, gazebo_ros_openni_kinect.cpp looks like it should be producing color. Are there plans to get the kinect gazebo working in color soon?

We got the tod_training package working, according to the BaseCreation tutorial. I noticed a program called "model_mesher," whose code looks like it is supposed to output a ".ply" (mesh) file from a bag file with a point cloud as input. Unfortunately, there were assertion errors:

/tmp/buildd/ros-diamondback-perception-pcl-0.10.0/debian/ros-diamondback-perception-pcl/opt/ros/diamondback/stacks/perception_pcl/flann/include/flann/algorithms/kdtree_single_index.h:151: void flann::KDTreeSingleIndex<distance>::BoundingBox::computeFromData(const flann::Matrix<typename distance::elementtype="">&) [with Distance = flann::L2_Simple<float>]: Assertion `data.rows>0' failed.

I am assuming this is still under development, since it was updated very recently. It doesn't look like it is being used in the OpenCV solutions in perception challenge. What are the plans for the model_mesher program? Is it still under active development? Should we try to improve it ourselves?

Thanks

I have seen several different solutions to 6D pose estimation in ROS/opencv:

checkerboard_detector

checkerboard_detector2

checkerboard_pose_estimation

datamatrix

fiducial

datamatrix seems less accurate than the checkerboard approach, but the checkerboard packages seem to no longer be supported. The fiducial package seems new (used in the tod_training contest), but there is no documentation.

Is willow/openCV settling on the fiducial package? I'd like to choose the one that works best right now and also won't disappear any time soon... Thanks!

The ecto framework http://ecto.willowgarage.com/ looks interesting. Do the developers plan to allow nodes to be located on different physical computers connected by a network?

We have used fiducial markers from tod_training for collecting training data and pose estimation and it worked fine so far (using object_recognition in SVN 50321).

I am working on an application that requires an exact correspondence between the image and depth information for a kinect in simulation and in real life. Using images and depth information from the turtlebot_simulator in gazebo and a real kinect, there seems to be a roughly linear relationship between the simulated kinect and the real kinect, but they are off by several centimeters. The following equation is in meters:

z_sim = 1.0816*z_real - 0.0123

To get this equation, I first measured a known object and constructed an exact mesh of that object, and then moved the simulated mesh in front of the turtlebot kinect, trying to align the real object in the real kinect image to look exactly the same as the simulated image from the simulated kinect by adjusting the location of the real object. Once I had a good alignment, I recorded the actual depth verse the simulated depth. After repeating this several times for different depths, I fit a line to the data, and got the above equation.

Why does this difference exist? Can it be fixed to be more accurate?

Thanks

While this may be an issue with OpenCV, just a heads up that the ROS team has stated that unstable will be very unstable during this early stage of development as low-level dependencies get a major overhaul for the upcoming Fuerte release. It is probably better to install the Electric release for most users.

When I try to install unstable-desktop-full, by running

sudo apt-get install ros-unstable-desktop-full

on Ubuntu 10.04 (Lucid Lynx), I get the following output:

The following packages have unmet dependencies:
  ros-unstable-desktop-full: Depends: ros-unstable-vision-opencv (= 1.6.1-s1313731218~lucid) but it is not going to be installed
                         Depends: ros-unstable-image-pipeline (= 1.5.2-s1313736766~lucid) but it is not going to be installed
                         Depends: ros-unstable-image-transport-plugins (= 1.4.1-s1313731295~lucid) but it is not going to be installed
E: Broken packages

Any ideas why?

It actually is in color, it just turns the pixels that are beyond the maximum depth of the kinect sensor to a gray color.