warning: cudaErrorIllegalAddress an illegal memory access was encountered

[sczgzzzj]

Thank you for your work, using GPU can greatly improve the speed, but the bag running larger scenes will often prompt insufficient memory, I am 3060 12G graphics card, can you save the optimized results to the hard disk at the same time, to avoid the problem of insufficient memory, thank you

[koide3]

There are a few parameters that can largely reduce the GPU memory usage. Please try setting them as follows.

• "sub_mapping/max_keyframe_overlap" : Decreasing this (e.g., to 0.3 ~ 0.4) makes submap creation less frequently.
• "sub_mapping/submap_target_num_points" : Decreasing this reduces the number of points in each submap. Usually, it's safe to reduce this to 10000 ~ 20000.
• "global_mapping/min_implicit_loop_overlap" : Increasing this (e.g., to 0.4 ~ 0.5) prevents creation of too much matching factors.

Storing intermediate mapping data in a disk would be hard to do properly (in particular with GPU memory). While we may consider it in future, we will not do it soon.

[sczgzzzj]

According to your above settings, it really greatly reduces the memory consumption of GPU. Thank you very much. Can the offline_viewer of GLIM now add the function of Multiple map merging like interactive_slam? This function is very useful. It can merge several bags calculated separately. Or can interactive_slam support GLIM output packets

[sczgzzzj]

In addition, I would like to ask you if you can add 2 graphics cards and calculate and increase the memory at the same time for calculating large scenes.Can I configure it in your source file？

[koide3]

Can the offline_viewer of GLIM now add the function of Multiple map merging like interactive_slam?

It is in a future plan, but for now, no support on merging multi maps.

I would like to ask you if you can add 2 graphics cards and calculate and increase the memory at the same time for calculating large scenes.

I expect drastic changes are required. Utilizing multi-GPUs for speeding up is relatively easy, but expanding the memory space is not. I think introducing some out-of-core GPU memory access mechanism should be a more realistic solution.