HOTFormerLoc: Hierarchical Octree Transformer for Versatile Lidar Place Recognition Across Ground and Aerial Views
- [2025-03-26] Training and evaluation code released. CS-Wild-Places dataset released.
This is the official repository for the paper:
HOTFormerLoc: Hierarchical Octree Transformer for Versatile Lidar Place Recognition Across Ground and Aerial Views, CVPR 2025 by Ethan Griffiths, Maryam Haghighat, Simon Denman, Clinton Fookes, and Milad Ramezani
[Website] [arXiv] [CS-Wild-Places Dataset] [CS-Wild-Places README]
HOTFormerLoc Architecture
We present HOTFormerLoc, a novel and versatile Hierarchical Octree-based TransFormer for large-scale 3D place recognition. We propose an octree-based multi-scale attention mechanism that captures spatial and semantic features across granularities, making it suitable for both ground-to-ground and ground-to-aerial scenarios across urban and forest environments.
In addition, we introduce our novel dataset: CS-Wild-Places, a 3D cross-source dataset featuring point cloud data from aerial and ground lidar scans captured in four dense forests. Point clouds in CS-Wild-Places contain representational gaps and distinctive attributes such as varying point densities and noise patterns, making it a challenging benchmark for cross-view localisation in the wild.
CS-Wild-Places dataset. (Top row) birds eye view of aerial global maps from all four forests.
(Bottom row) sample ground and aerial submap from each forest.
Our results demonstrate that HOTFormerLoc achieves a top-1 average recall improvement of 5.5% – 11.5% on the CS-Wild-Places benchmark. Furthermore, it consistently outperforms SOTA 3D place recognition methods, with an average performance gain of 4.9% on well established urban and forest datasets.
If you find this work useful, please consider citing:
@InProceedings{HOTFormerLoc,
author = {Griffiths, Ethan and Haghighat, Maryam and Denman, Simon and Fookes, Clinton and Ramezani, Milad},
title = {{HOTFormerLoc}: {Hierarchical Octree Transformer} for {Versatile Lidar Place Recognition Across Ground} and {Aerial Views}},
booktitle = {2025 {IEEE}/{CVF Conference} on {Computer Vision} and {Pattern Recognition} ({CVPR})},
year = {2025},
month = {June},
}
Code was tested using Python 3.11 with PyTorch 2.1.1 and CUDA 12.1 on a Linux system. We use conda to manage dependencies (although we recommend mamba for a much faster install).
# Note: replace 'mamba' with 'conda' if using a vanilla conda install
mamba create -n hotformerloc python=3.11 -c conda-forge
mamba activate hotformerloc
mamba install 'numpy<2.0' -c conda-forge
mamba install pytorch==2.1.1 torchvision==0.16.1 pytorch-cuda=12.1 -c pytorch -c nvidia -c conda-forge
pip install -r requirements.txt
pip install libs/dwconv
Modify the PYTHONPATH environment variable to include the absolute path to the repository root folder (ensure this variable is set every time you open a new shell):
export PYTHONPATH=$PYTHONPATH:<path/to/HOTFormerLoc>
We train on the Wild-Places dataset introduced in Wild-Places: A Large-Scale Dataset for Lidar Place Recognition in Unstructured Natural Environments (link).
Download the dataset here, and place or symlink the data in data/wild_places.
Run the following to fix the broken timestamps in the poses files:
cd datasets/WildPlaces
python fix_broken_timestamps.py \
--root '../../data/wild_places/data/' \
--csv_filename 'poses.csv' \
--csv_savename 'poses_fixed.csv' \
--cloud_folder 'Clouds'
python fix_broken_timestamps.py \
--root '../../data/wild_places/data/' \
--csv_filename 'poses_aligned.csv' \
--csv_savename 'poses_aligned_fixed.csv' \
--cloud_folder 'Clouds_downsampled'
Before network training or evaluation, run the below code to generate pickles with positive and negative point clouds for each anchor point cloud:
cd datasets/WildPlaces
python generate_training_tuples.py \
--root '../../data/wild_places/data/'
python generate_test_sets.py \
--root '../../data/wild_places/data/'
We train on our novel CS-Wild-Places dataset, introduced in further detail in our paper. CS-Wild-Places is built upon the ground traversals introduced by Wild-Places, so it is required to download the Wild-Places dataset alongside our data following the instructions in the above section (generating train/test pickles for Wild-Places is not required for CS-Wild-Places, so this step can be skipped). Note that the full Wild-Places dataset must be downloaded as our post-processing utilises the full resolution submaps.
Download our dataset from CSIRO's data access portal, and place or symlink the data in data/CS-Wild-Places (this should point to the top-level directory, with the data/ and metadata/ subdirectories). Note that our experiments only require the post-processed submaps (folder postproc_voxel_0.80m_rmground_normalised), so you can ignore the raw submaps if space is an issue. Check the README for further information and installation instructions for CS-Wild-Places.
Assuming you have followed the above instructions to setup Wild-Places, you can use the below command to post-process the Wild-Places ground submaps into the format required for CS-Wild-Places (set num_workers to a sensible number for your system). Note that this may take several hours depending on your CPU:
cd datasets/CSWildPlaces
python postprocess_wildplaces_ground.py \
--root '../../data/wild_places/data/' \
--cswildplaces_save_dir '../../data/CS-Wild-Places/data/CS-Wild-Places/postproc_voxel_0.80m_rmground_normalised' \
--remove_ground \
--downsample \
--downsample_type 'voxel' \
--voxel_size 0.8 \
--normalise \
--num_workers XX \
--verbose
Note that this script will generate the submaps used for the results reported in the paper, i.e. voxel downsampled, ground points removed, and normalised. We also provide a set of unnormalised submaps for convenience, and the corresponding Wild-Places ground submaps can be generated by omitting the --normalise option, and by setting --cswildplaces_save_dir to '../../data/CS-Wild-Places/data/CS-Wild-Places/postproc_voxel_0.80m_rmground'.
Before network training or evaluation, run the below code to generate pickles with positive and negative point clouds for each anchor point cloud:
cd datasets/CSWildPlaces
python generate_train_test_tuples.py \
--root '../../data/CS-Wild-Places/data/CS-Wild-Places/postproc_voxel_0.80m_rmground_normalised/' \
--eval_thresh '30' \
--pos_thresh '15' \
--neg_thresh '60' \
--buffer_thresh '30' \
--v2_only
Note that training and evaluation pickles are saved to the directory specified in --root by default.
We train on the CS-Campus3D dataset introduced in CrossLoc3D: Aerial-Ground Cross-Source 3D Place Recognition (link).
Download the dataset here, and place or symlink the data in data/benchmark_datasets_cs_campus3d.
Run the below commands to convert the CS_Campus3D train and test pickles into a suitable format for use with HOTFormerLoc.
cd datasets/CSCampus3D
python save_queries_HOTFormerLoc_format.py
We trained on a subset of Oxford RobotCar and the In-house (U.S., R.A., B.D.) datasets introduced in PointNetVLAD: Deep Point Cloud Based Retrieval for Large-Scale Place Recognition (link). There are two training datasets:
- Baseline Dataset - consists of a training subset of Oxford RobotCar
- Refined Dataset - consists of training subset of Oxford RobotCar and training subset of In-house
We report results on the Baseline set in the paper.
For dataset description see the PointNetVLAD paper or github repository (link).
You can download the dataset from
here
(alternative link), then place or symlink the data in data/benchmark_datasets.
Before network training or evaluation, run the below code to generate pickles with positive and negative point clouds for each anchor point cloud.
cd datasets/pointnetvlad
# Generate training tuples for the Baseline Dataset
python generate_training_tuples_baseline.py --dataset_root '../../data/benchmark_datasets'
# (Optionally) Generate training tuples for the Refined Dataset
python generate_training_tuples_refine.py --dataset_root '../../data/benchmark_datasets'
# Generate evaluation tuples
python generate_test_sets.py --dataset_root '../../data/benchmark_datasets'
To train HOTFormerLoc, download the datasets and generate training pickles as described above for any dataset you wish to train on.
The configuration files for each dataset can be found in config/.
Set the dataset_folder parameter to the dataset root folder (only necessary if you have issues with the default relative path).
If running out of GPU memory, decrease batch_split_size and val_batch_size parameter value. If running out of RAM, you may need to decrease the batch_size parameter or try reducing num_workers to 1, but note that a smaller batch size may slightly reduce performance. We use wandb for logging by default, but this can be disabled in the config.
To train the network, run:
cd training
# To train HOTFormerLoc on CS-Wild-Places
python train.py --config ../config/config_cs-wild-places.txt --model_config ../models/hotformerloc_cs-wild-places_cfg.txt
# To train HOTFormerLoc on Wild-Places
python train.py --config ../config/config_wild-places.txt --model_config ../models/hotformerloc_wild-places_cfg.txt
# To train HOTFormerLoc on CS-Campus3D
python train.py --config ../config/config_cs-campus3d.txt --model_config ../models/hotformerloc_cs-campus3d_cfg.txt
# To train HOTFormerLoc on Oxford RobotCar
python train.py --config ../config/config_oxford.txt --model_config ../models/hotformerloc_oxford_cfg.txt
If training on a SLURM cluster, we provide the submitit_train_job_single_node.py script to automate training job submission, with support for automatic checkpointing and resubmission on job timeout. Make sure to set job parameters appropriately for your cluster.
Pre-trained weights for HOTFormerLoc and other experiments can be downloaded and placed in the weights directory. You can download them individually below, or download and extract all from this link.
| Model | Dataset | Weights Download |
|---|---|---|
| HOTFormerLoc | CS-Wild-Places | hotformerloc_cs-wild-places.pth |
| HOTFormerLoc | CS-Campus3D | hotformerloc_cs-campus3D.pth |
| HOTFormerLoc | Wild-Places | hotformerloc_wild-places.pth |
| HOTFormerLoc | Oxford RobotCar | hotformerloc_oxford.pth |
| MinkLoc3Dv2 | CS-Wild-Places | minkloc3dv2_cs-wild-places.pth |
| CrossLoc3D | CS-Wild-Places | crossloc3d_cs-wild-places.pth |
| LoGG3D-Net | CS-Wild-Places | logg3dnet_cs-wild-places.pth |
To evaluate the pretrained models run the following commands:
cd eval
# To evaluate HOTFormerLoc trained on CS-Wild-Places
python pnv_evaluate.py --config ../config/config_cs-wild-places.txt --model_config ../models/hotformerloc_cs-wild-places_cfg.txt --weights ../weights/hotformerloc_cs-wild-places.pth
# To evaluate HOTFormerLoc trained on Wild-Places
python pnv_evaluate.py --config ../config/config_wild-places.txt --model_config ../models/hotformerloc_wild-places_cfg.txt --weights ../weights/hotformerloc_wild-places.pth
# To evaluate HOTFormerLoc trained on CS-Campus3D
python pnv_evaluate.py --config ../config/config_cs-campus3d.txt --model_config ../models/hotformerloc_cs-campus3d_cfg.txt --weights ../weights/hotformerloc_cs-campus3d.pth
# To evaluate HOTFormerLoc trained on Oxford RobotCar
python pnv_evaluate.py --config ../config/config_oxford.txt --model_config ../models/hotformerloc_oxford_cfg.txt --weights ../weights/hotformerloc_oxford.pth
Below are the results for all evaluated models on CS-Wild-Places:
Comparison of SOTA on CS-Wild-Places Baseline evaluation set.
Comparison of SOTA on CS-Wild-Places Unseen evaluation set.
See the paper for full results and comparison with SOTA on all datasets.
Special thanks to the authors of MinkLoc3Dv2 and OctFormer for their excellent code, which formed the foundation of this codebase. We would also like to thank the authors of Wild-Places for their fantastic dataset which serves as the base that CS-Wild-Places is built upon.