Request for Training/Testing Codes of MinkLoc3Dv2, CrossLoc3D, LoGG3D-Net on CS-Wild-Places

[typist2001]

Hello, could you please provide the training and testing codes for MinkLoc3Dv2, CrossLoc3D, and LoGG3D-Net on the CS-Wild-Places dataset? I would like to know how to test the results of the model weights you provided.

[Grifent]

Hi. I will try to provide a minimal example in due course.
In the meantime, we use the same train/eval tuple format as MinkLoc3Dv2, so you should be able to use the CS-Wild-Places evaluation pickles with MinkLoc3Dv2's evaluation script. You will just need to point to the correct query/database files in lines 24-28 and copy over the point cloud loader (and create a new config file based on config_baseline_v2.txt with the correct path to CS-Wild-Places).
For LoGG3D-Net, you can follow the general instructions from this issue and here to run LoGG3D-Net on a custom dataset and pre-compute the embeddings (make sure that LoGG3D-Net feature_dim is set to 32). You can then use the Wild-Places inter-sequence.py evaluation script, just make sure you pass the correct paths/names for the CLI args.
CrossLoc3D is a little trickier as it requires downsampling point clouds to 4096 points (which can be done with postprocess_submaps.py --remove_ground --downsample --downsample_target 4096 --downsample_type random --normalise options), and then creating a new cfg file for each forest of CS-Wild-Places + adding the correct dataloader, etc.