jac99 / minkloc3dv2 Goto Github PK

The file "evaluate.py" is missing from the eval directory.

Dear author,

I am using the learning-based descriptor for place recognition in a SLAM task.
However, the top 1 candidate is always returned though there is in fact no revisits.
So I guess we need a threshold to discard negative place recognitions.
But how do we determine this threshold?
Or do you have a better alternative to tell positive and negative place recognitions apart?
Note that we are working with a large scale dataset, the lidar-based odometry fastlio2 may drift severely say 40 meters.

Best,

Hi, thanks for your amazing work!

I tried running train.py but encountered a problem with too much RAM usage. There are about 40GB RAM on my machine, but even before the first epoch finished, it already ran out. I decreased the batch_split_size as suggested, but the still ran into the same problem.

I checked the memroy usage of each part inside do_train(), seems like the problem is with make_dataloaders() in dataset_utils.py , which consumes a lot of memory.

Could I please know how much RAM train.py actually takes, and if there are possible solutions to my problem? Thanks in advance!

how about the pointcloud generated by depth image? such as real sense, kinect .... ?
thanks

Hi, thanks for your great job!
I am gonna to test your work in our project. We have done some investigation before. In some lidar-based place recognition method, lidar with 360 degree data with at least 36 lines is recommended. How about the adaptability to different lidar sensor of your method?
Thank your for your attention and keep waiting for kind response!

Dear Author,
I'm sorry to bother you, I have met this problem when I ran to 35 epoches:
" File "~/venv38/lib/python3.8/site-packages/MinkowskiEngine-0.5.4-py3.8-linux-x86_64.egg/MinkowskiEngine/MinkowskiConvolution.py", line 72, in forward
return fw_fn(
MemoryError: std::bad_alloc: cudaErrorMemoryAllocation: out of memory"
The version of MinkowskiEngine is 0.5.4, but the version of cuda is 11.6, I wonder to know if this problem is caused by the version of cuda, I couldn't change the version of cuda casully for it may cause trouble to other users.
I'm looking forward to your reply!
Best Wishes!

I have read your paper and run the code you provide on the Github. It is a great work.

I would like to use MinkLoc3Dv2 for point cloud retrieval by using the global descriptor. in the work of figure 3. You used the query point clouds to retrieve the corresponding point clouds.

I would like to know Is gnss data necessary for this process?

Dear Sir,

I am on ubuntu 20. As your readme recommend cuda 10.2. I tried to install it.
However, the available cuda from apt-cache is cuda 11.x as shown here.
Also, I see that people are advising not to use cuda 10.2 on ubuntu 20.04.
Since the repo is two years old, do you still recommend using cuda 10.2 on ubuntu 20?

If this is the case, I may proceed with the guide here. Do you think this guide is viable?

Thanks for your wonderful job! I want to train/fine tune it on my own dataset, could you please give me some suggestions about how to modify the code? Thanks again!