WildCross: A Cross-Modal Large Scale Benchmark for Place Recognition and Metric Depth Estimation in Natural Environments

Accepted at IEEE ICRA 2026

Joshua Knights^1,2 , Joseph Reid¹ , Kaushik Roy¹ , David Hall¹ , Mark Cox¹ , Peyman Moghadam^1,2

¹CSIRO Robotics   ²Queensland University of Technology

This repository contains supporting scripts for downloading, training and evaluating techniques on WildCross, a large-scale multi-modal benchmark for place recognition and metric depth estimation in natural environments accepted at IEEE ICRA2026.

News

• FEBRUARY 2026: Paper, project page, code, models, demo, and benchmark are all released.
• JANUARY 2026: WildCross Paper is accepted to IEEE ICRA 2026.

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.

Usage

Data loading and visualisation scripts

Instructions on how to load each of the data modalities (RGB Image, Depth Image, Surface Normal Image, 3D Submap) can be found in example_loaders.py, and a script to visualise the depth image using the palette employed in our visualisations can be found in visualise_depth_images.py

Training and Evaluation

We provide code for training and evaluation on the WildCross dataset for the tasks of visual place recognition (VPR), cross-modal place recognition (CMPR) and metric depth estimation, which can be found in their respectives subfolders inside the benchmarking folder. For more detailed instructions for setting up and running these benchmarks, consult the documentation inside the respective subfolders for each task. For LiDAR place recognition (LPR) code for training and evaluation can be found on a new WildCross_subsets branch of the original Wild-Places repository. Information for getting to the branch can be found inside the benchmarking/LPR folder.

Checkpoints

We provide checkpoints for the fine-tuned models used to produce the results in this publication, as well as the urban pre-trained models provided by the original authors for each of the benchmarked methods where applicable. See the sub-folder for each benchmarked task to find the respective download links to the relevant checkpoints for that task. You can also find all weights on the WildCross HuggingFace page

Acknowledgements

We are grateful to the authors and open source maintainers for NetVlad, MixVPR, SALAD, BOQ, LIP-Loc and DepthAnythingV2, whose implementations form the basis of the training and evaluation code presented in this repository. We would also like to thank the author of the VPR-methods-evaluation repository, which provided an excellent starting point for the development of the evaluation scripts in this repository.

Citation

If you find this repository useful or use the WildCross dataset in your work, please cite the paper using the following:

@inproceedings{wildcross2026,
  title={{WildCross: A Cross-Modal Large Scale Benchmark for Place Recognition and Metric Depth Estimation in Natural Environments}},
  author={Joshua Knights, Joseph Reid, Kaushik Roy, David Hall, Mark Cox, Peyman Moghadam},
  booktitle={Proceedings-IEEE International Conference on Robotics and Automation},
  pages={},
  year={2026}
}

And the original WildPlaces paper:

@inproceedings{knights2023wildplaces,
  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={Proceedings-IEEE International Conference on Robotics and Automation},
  pages={11322--11328},
  year={2023}
}