stlucia datasets can not get the correct map

[orange2xiaoyu]

Hi, I download your code and run it using irat_red datasets and stlucia datasets, but I can not get the correct map that same as paper, the map as following. I have not modified any parameters in config file.

[orange2xiaoyu]

I changed a computer to run ratslam and got the following map:

Which parameter should I modify to get the results in the article？

[FelixSteinbauer]

Hi, I had the same problem with the repo.
But I do not think its actually an issue of the repository itself but with ROS feeding the data to fast.

When playing the data with normal speed (rosbag play stlucia_2007.bag) I get the following result.

When feeding the data with half the speed (rosbag play stlucia_2007.bag --rate=0.5) I get a proper topographical map:

That the result is squeezed vertically (in comparison to Fig.9 in the paper) is normal I think and just related to the rectangular visualization we use.

Speculation about the underlying issue
We assume that the issue originates from the (linearly?, see Fig. 14) growing amount of vision templates the system learns over time. In the beginning, each visual scene can be compared to existing templates before the next frame is fed into the system. But after a certain point, the re-localization algorithm cannot finish in time resulting in no further loops to be closed and therefore no correction of further map/experience points.

However, I am no ROS or RatSLAM expert, so I can just speculate why decreasing the playback speed fixes this issue.

[orange2xiaoyu]

Thank you for your reply, I try to feed the data with half speed, got the following topographical map:

And I want to know the configuration of your computer running this code?

[FelixSteinbauer]

Did you generate the latest graphic on your original computer or the second one you mentioned in this thread?
If it was on the first one; I would say there is an improvement and you should maybe try even less speed.
If that was on the second machine, I don't see improvement. I cannot help you either in that case. Decreasing the speed worked at least for me and one colleague.

I am running this fork of the repository on an Intel i7 8 core (3,4 GHz) machine with 20GB RAM and a NVIDIA GTX 1660.
I have Ubuntu 20.04 with ROS noetic, opencv 4.2 and irrlicht 1.8.

[orange2xiaoyu]

I run this code on the second machine.

I am trying to run this fork you used, and I do not modify any parameters. I get the following topographical map(only part of data have been fed), and the loops should be closed in the red rectangle(the data with half speed).

I have ubuntu 18.04 with ROS melodic ,opencv3.2 and irrlicht 1.8, and my machine with the following configuration:

[SeaEastXu]

I have the same quesiton，do you solve it and reproduce the correct map in paper?

[orange2xiaoyu]

You can try to feed the data with low speed, I get the correct map using rate=1Hz.

[SeaEastXu]

1Hz?
Do you mean that the rate factor is 0.01 and clock time at frequency is 1?

[orange2xiaoyu]

You can feed the data with half speed if your machine configuration is relatively high.