search

AndreiBarsan / DynSLAM

Master's Thesis on Simultaneous Localization and Mapping in dynamic environments. Separately reconstructs both the static environment and the dynamic objects from it, such as cars.
BSD 3-Clause "New" or "Revised" License
576 stars 179 forks source link

Compared to the original InfiniTAM, does this implemention need parameter adjustment? #60

Closed MicroYY closed 5 years ago

MicroYY commented 5 years ago

I am wondering that since the original InfiniTAM works indoor and DynSLAM works outdoor, how to adjust the parameters so that InfiniTAM can work outdoor? The depth information is quite different in these two situations.

AndreiBarsan commented 5 years ago

This depends on the sensor you are using, how your stereo rig is configured, etc. Also, in many outdoor scenarios the motion tends to be faster than in indoor ones (e.g., in KITTI you can have very fast motion but only 10Hz inputs), so you'd need to use a different camera tracking method.

Do you have any details about the context in which you wish to make this work? Do you want to run DynSLAM on indoor data?

MicroYY commented 5 years ago

Thanks for your reply! I want to use a stereo camera which outputs depth map in real time for outdoor reconstruction. I employ VINS-MONO as the visual odometry, so the camera pose is well got. Now the InfiniTAM can reconstruct indoor scenarios quite well, but when it comes to outdoor objects, the result is not that satisfactory. My stereo camera outputs depth up to 20m, and obviously the original InfiniTAM cannot fuse depth that far. I knwo there are several parameters affecting the reconstrucion quality. The code defines a class called sceneParams which wraps the parameters related to the scene representation. I tried several combinations of these parameters but cannot achieve a satisfactory result. So how do DynSLAM adjust the parameters so that it can work outdoor? Do you have a set of parameters? If the parameters are suitable, maybe the problem comes from depth map.

AndreiBarsan commented 5 years ago

Yes, it's likely an issue with how the depth is computed. Have you tried tweaking the calibration parameters, e.g., the stuff from the sample file here: https://github.com/AndreiBarsan/DynSLAM/blob/a7e885205de9bf8e40527c834398c63ad2f90641/itm-sample-calib-from-kitti-odometry-sequence-06.txt ?

This calibration is read here in the code: https://github.com/AndreiBarsan/DynSLAM/blob/a7e885205de9bf8e40527c834398c63ad2f90641/src/DynSLAM/DynSLAMGUI.cpp#L1162

The last line in the calibration file is very important. This thread could be helpful in explaining it: https://github.com/victorprad/InfiniTAM/issues/130

Meanwhile, can you describe what is happening? Does the reconstruction just not show anything?

AndreiBarsan commented 5 years ago

Closing this due to lack of activity. Please don't hesitate to reach out if you still have questions/problems!