The site is read-only. Please transition to use Robotics Stack Exchange
| ROS Resources: Documentation | Support | Discussion Forum | Index | Service Status | ros @ Robotics Stack Exchange |
 | 1 | initial version |
So I gave it a go and used tf::TransformListener::transfromPointCloud and it turned out to be much much faster (CPU usage went from hogging a core to 10%). I haven't tried pcl::transformPointCloud.
| | 2 | No.2 Revision |
So I gave it a go and used tf::TransformListener::transfromPointCloud and it turned out to be much much faster (CPU usage went from hogging a core to 10%). I haven't tried pcl::transformPointCloud.pcl::transformPointCloud yet, but I guess I will, because that would be interesting to know which one is faster.
I am a bit puzzled though. I am copying my points to a point cloud, then they get copied to a ublas matrix. This matrix is then multiplied by the transform. Then points are copied back to a point cloud and back to my data format. So first of all there are 2 copy operations (but in my own implementation I also had to copy to a btVector3 back and forth ...). Then the number of mathematical operations is still the same, isn't it: the 600 points still have to be multiplied by the transform. So either ublas is exceptionally well optimized, or my implementation was really crappy.
| | 3 | updated after migrating from diamondback to fuerte |
So I gave it a go and used tf::TransformListener::transfromPointCloud and it turned out to be much much faster (CPU usage went from hogging a core to 10%). I haven't tried pcl::transformPointCloud yet, but I guess I will, because that would be interesting to know which one is faster.
I am a bit puzzled though. I am copying my points to a point cloud, then they get copied to a ublas matrix. This matrix is then multiplied by the transform. Then points are copied back to a point cloud and back to my data format. So first of all there are 2 copy operations (but in my own implementation I also had to copy to a btVector3 back and forth ...). Then the number of mathematical operations is still the same, isn't it: the 600 points still have to be multiplied by the transform. So either ublas is exceptionally well optimized, or my implementation was really crappy.
EDIT: Since then, I have switched from diamondback to fuerte. I noticed that the implementation of tf::TransformListener::transfromPointCloud has changed, starting from electric. It is now transforming all the points one by one. See the following API links for a comparison:
I haven't thoroughly timed them, but it seems that it's still doing a good job at it.
ROS Answers is licensed under Creative Commons Attribution 3.0 Content on this site is licensed under a Creative Commons Attribution Share Alike 3.0 license.
