search

felixendres / rgbdslam_v2

RGB-D SLAM for ROS
GNU General Public License v3.0
943 stars 400 forks source link

How to refining the registration by ICP? #68

Open blueeaglex opened 7 years ago

blueeaglex commented 7 years ago

The only information I can get is "GICP Generalized ICP can be (de)activated for refining the registration. For more information see http://stanford.edu/~avsegal/generalized_icp.html" But the url is not accessable. How can I get the registered pointcloud sequence rather than a merged large single pointcloud(The quicksave button)?

felixendres commented 7 years ago

Using ICP mainly was an experiment by a co-author. Since it didn't improve results consistently in acceptable time it hasn't been used for a/usr/lib/libcxsparse.so.3.1.2 loong time and I do not recommend using it. For some results, see Tabelle 2 in this (German) publication.

For the sequence of clouds, there's the option "Save Point Cloud Node-Wise..." in the "Save" menu. Or you can save as a bag file, then you get the original clouds plus the transformations in tf messages.

blueeaglex commented 7 years ago

@felixendres Thank you very much Although I can not read German, I can see the results. Yes,the result is not reasonable. I also tested ICP registration 1 years ago. It's not as good as the rumors. But by combining point-to-point ICP，point-to-plane ICP and lum, It could be much better. Of course, this is just a simple combination, and I ignored the cost of time. In other words，I don't care about time， I just want good results. But even then, it was still not good enough. rgbdslam_v2 did much better. Well. Now I can get the sequence, and I will be able to restart my experiment on a sequence of registered point cloud. I'll share the results if I get something reasonable. Thanks again.