Wild-Places: A Large-Scale Dataset for Lidar Place Recognition in Unstructured Natural Environments

This repository contains the code implementation used in the paper Wild-Places: A Large-Scale Dataset for Lidar Place Recognition in Unstructured Natural Environments, which has been published at ICRA2023.

If you find this dataset helpful for your research, please cite our paper using the following reference:

@inproceedings{2023wildplaces,
  title={Wild-places: A large-scale dataset for lidar place recognition in unstructured natural environments},
  author={Knights, Joshua and Vidanapathirana, Kavisha and Ramezani, Milad and Sridharan, Sridha and Fookes, Clinton and Moghadam, Peyman},
  booktitle={2023 IEEE international conference on robotics and automation (ICRA)},
  pages={11322--11328},
  year={2023},
  organization={IEEE}
}

Contents

1. Updates
2. Download Instructions
3. Benchmarking
   • Checkpoints
   • Performance
4. Scripts
   • Loading Point Clouds
   • Training
   • Evaluation
5. Thanks

Updates

• Oct 2022 Wild-Places v1.0 Uploaded
• Jan 2023 Wild-Places is accepted to ICRA2023!
• Jan 2023 Wild-Places v2.0 Uploaded. This update to the dataset integrates GPS into the SLAM solution to alleviate vertical drift in the larger loops of the traversal in both environments. NOTE: Sequence K-04 is currently unavailable for v2.0 due to a failed loop closure in the ground truth. We are currently working on remedying this, and will release the sequence as soon this issue is rectified.
• Feb 2025 Fix the broken timestamps in the poses files.
• Nov 2025 Wild-Places v3.0 Uploaded. This update to the dataset includes:
  • Updated point clouds / trajectories using the latest version of WildCat SLAM to bring the dataset in line with the pointclouds available in the WildScenes and WildCross datasets
  • Updated benchmarking results and instructions for training on LoGG3D-Net and MinkLoc3Dv2
  • Updated dataset and repository file structure
  • Intensity channel has been added to submaps
  • Providing WildCross branch LPR splits

Download Instructions

Our dataset can be downloaded through The CSIRO Data Access Portal. Detailed instructions for downloading the dataset can be found in the README file provided on the data access portal page.

Benchmarking

Here we provided pre-trained checkpoints and results for benchmarking several state-of-the-art LPR methods on the Wild-Places dataset.

Update Nov. 2025: With the release of Wild-Places v3.0, we have also re-run training for two state-of-the-art methods (LoGG3D-Net, MinkLoc3Dv2) on the Wild-Places dataset using expanded batch sizes to provide new training checkpoints which better reflect the capabilities of recent state-of-the-art GPUs. We provide checkpoints and benchmarked results for both the recently trained models (2025) and the checkpoints released with the ICRA2023 paper.

Pre-trained Model Checkpoints

Release	Model	Checkpoint
ICRA2023	TransLoc3D	🤗 HuggingFace
ICRA2023	MinkLoc3DV2	🤗 HuggingFace
ICRA2023	LoGG3D-Net	🤗 HuggingFace
2025 Re-Training	MinkLoc3DV2	🤗 HuggingFace
2025 Re-Training	LoGG3D-Net	🤗 HuggingFace

Performance

Scripts

Loading Point Clouds

A code snippet to load a pointcloud file from our dataset can be found in eval/load_pointcloud.py

Training

We provide instructions for how to add Wild-Places as a training dataset for two state-of-the-art LPR methods: LoGG3D-Net and MinkLoc3Dv2. For more detailed instructions, please consult the README.md files in training/LoGG3D-Net and training/MinkLoc3Dv2.

Evaluation

We provide generic evaluation code for evaluating performance on the Wild-Places dataset for both the inter and intra-sequence testing scenarios, as well as an implementation of ScanContext. For more details, please see the README.md file in the eval folder.

Thanks

Special thanks to the authors of the PointNetVLAD and MinkLoc3D, whose excellent code was used as a basis for the generation and evaluation scripts used in this repository.