😉The code is almost there. Don't ask, we don't know either. Good luck!😉 [kidding]
source code of our CVPR'25 paper ForestLPR: LiDAR Place Recognition in Forests Attentioning Multiple BEV Density Images
- ForestLPR: LiDAR Place Recognition in Forests Attentioning Multiple BEV Density Images
- CUDA 11.8
- Python 3.9.4
- PyTorch 2.0.1
We used Anaconda to setup a deep learning workspace that supports PyTorch. Run the following script to install the required packages.
conda env create -f environment.yml
conda activate forestlpr
git clone [email protected]:shenyanqing1105/ForestLPR-CVPR2025.git
cd ForestLPR-CVPR2025
pip install -r requirements.txt
conda deactivateWe use two public datasets and one own-collected dataset: Wild-Places, Botanic Garden Dataset, and ANYmal dataset. Additionally, in our work, we have processed the datasets into cropped point clouds Cloud_normalized, pickle files and BEV density images .npy files, which is more convenient to test the BEV-based methods.
For Botanic-code: MPTw, and ANYmal-code: 6hXL, we repurpose the dataset and use GT to register multiple frames to 1 submap.
For Wild-Places-code: s0u3, we provide the additional processed files, and the data should be merged based on the sequences. [sorry for missing some pickle files, we upload them directly to the repo :)]
Environment visualization
$_PATH_TO_DATASET/
├── anymal/ # Dataset
| ├── anymal-3/ # Environment
| │ └── 02/ # Sequence
| | ├── Clouds_downsampled/
| | ├── Clouds_normalized/
| | └── poses_aligned.csv
| └── testing/
| ├── 3-02.pickle
| └── .npy files
├── botanic/ # Dataset
| ├── 1005-03/ # Environment
| │ └── 02/ # Sequence
| | ├── Clouds_downsampled/
| | ├── Clouds_normalized/
| | └── poses_aligned.csv
| ├── 1005-06/
| └── ...
| └── testing/
| ├── 5-03.pickle
| ├── 5-06.pickle
| └── .npy files
└── Wild-Places/
├── Karawatha/
| ├── K-01/
| | ├── Clouds_downsampled/
| | └── poses_aligned.csv
| ├── K-02/
| | ├── Clouds_downsampled/
| | └── poses_aligned.csv
| ├── K-03/
| | ├── Clouds_downsampled/
| | └── poses_aligned.csv
| └── K-04/
| ├── Clouds_downsampled/
| └── poses_aligned.csv
├── testing/
| ├── V-03.pickle
| ├── V-04.pickle
| ├── K-03.pickle
| ├── V-03.pickle
| └── .npy files
├── training/
| ├── training_wild-places.pickle
| ├── testing_wild-places.pickle
| └── .npy files
└── Venman/
├── V-01/
├── V-02/
├── V-03/
└── V-04/
$HOME/ForestLPR
├── dataset/ # Dataset
| ├── anymal-3/ # Environment
| │ └── 02/ # Sequence
| | ├── Clouds_downsampled/
| | ├── Clouds_normalized/
| | └── poses_aligned.csv
| └── testing/
| ├── 3-02.pickle
| └── npy files
├── models/
| ├── model/ # model files of ForestLPR
| | ├── Deit.py
| | ├── Deit_multi.py
| | └── ...
| └── bp/ # model files of BevPlace
├── pickle_files/
├── runs/ # checkpoints
├── scripts/
├── eval/ # evaluations for 2D baselines
| ├── mapclosure/
| | ├── mc_inter_sequence.py
| | ├── mc_intra_sequence.py
| | ├── mc_utils.py
| | ├── mc_eval_config_anymal.yaml
| | ├── mc_eval_config_botanic.yaml
| | └── mc_eval_config_wp.yaml
| ├── scan_context/
| | ├── sc_inter_sequence.py
| | ├── sc_intra_sequence.py
| | ├── sc_utils.py
| | ├── sc_eval_config_anymal.yaml
| | ├── sc_eval_config_botanic.yaml
| └── sc_eval_config_wp.yaml
| └── utils.py
├── generate_splits/
| ├── bevwp.py # dataloader of ForestLPR
| ├── ground_filter.py # point-cloud preprocessing: ground segmentation and height offset removal
| ├── merge.py # generating the submaps from original LiDAR scans
| ├── testing_sets.py
| ├── training_sets.py
| ├── util.py
| └── utils.py
├── load_pointcloud.py
└── overlap.py # compute the overlap between two scans using octree
├── eval.py # evaluations for forestLPR
├── eval_bp.py # evaluations for BevPLace
├── train.py
├── util.py
└── utils.py
$HOME/LoGG3D-Net
├── [original code]
├── scripts/
| ├── inter_sequence.py
| ├── intra_sequence.py
| ├── load_pointcloud.py
| ├── transloc_utils.py
| └── util.py
└── checkpoints
$HOME/MinkLoc3Dv2
├── [original code]
├── scripts
| ├── inter_sequence.py
| ├── intra_sequence.py
| ├── load_pointcloud.py
| ├── minkloc_utils.py
| └── util.py
└── checkpoints
$HOME/TransLoc3D
├── [original code]
├── scripts
| ├── inter_sequence.py
| ├── intra_sequence.py
| ├── load_pointcloud.py
| ├── logg3d_utils.py
| └── utils.py
└── checkpoints
First check dataset file path in bevwp.py
Run scripts/generate_splits/ground_filter.py, change <ROOT_PATH> and <OUT_PATH>
Including ground segmentation and height offeset removal
To generate the training splits run the following command:
python generate_splits/training_sets.py --dataset_root $_PATH_TO_DATASET --save_folder --$_SAVE_FOLDER_PATH
Where $_PATH_TO_DATASET is the path to the downloaded dataset, and $_SAVE_FOLDER_PATH is the path to the directory where the generated files will be saved.
To generate the testing splits run the following command:
python generate_splits/testing_sets.py --dataset_root $_PATH_TO_DATASET --save_folder --$_SAVE_FOLDER_PATH
This script will generate separate testing pickles for the inter-run and intra-run evaluation modes on each environment. The inter-run pickles will produce query and database files for each testing environment, while the intra-run pickles will produce a separate training pickle for each individual point cloud sequence.
For ANYmal and botanic, we also use the same way to generate pickle files, which are saved in dataset folders.
In scripts/generate_splits/bevwp.py, set TOP_Z_MAX and TOP_Z_MIN to crop the point cloud. TOP_Z_DIVISION refers to height interval of slices.
generate_bevs function is used to generate BEV images and save .npy files to the folder.
This script is also used for dataloader. When running the code first time, these files can be generated automatically. If there is already the .npy file, the dataset load function can skip the generation and directly load it from the file.
Run the following script to train and evaluate ForestLPR network. Specifically, it will train ForestLPR network and select a checkpoint that performs best Recall@5 on the validation set as the final model.
use Deit_multi as model file
python -W ignore train.py --img_size 480 480 --level 1 6 11 --agg gem --depth 12 --inputc 5 --features_dim 1024 --firsteval
use Deit as model file, train single-BEV version
python -W ignore train.py --img_size 480 480 --level 1 6 11 --agg gem --depth 12 --inputc 1 --features_dim 1024 --firsteval
Download the trained model on Wild-Places from checkpoints, put it into $HOME/ForestLPR/runs/forestlpr, and run the following script to evaluate it.
Change dataset file path in bevwp.py
Wild-Places inter-evaluation
python -W ignore eval.py --img_size 480 480 --cacheBatchSize 6 --resume ./runs/forestlpr/checkpoints/model_best.pth.tar --level 1 6 11 --dataset inter --subset KarawathaWild-Places intra-evaluation
python -W ignore eval.py --img_size 480 480 --cacheBatchSize 12 --resume ./runs/forestlpr/checkpoints/model_best.pth.tar --level 1 6 11 --dataset intra --subset V-03 --world_thresh 3 --time_thresh 600Botanic intra-evaluation
python -W ignore eval.py --img_size 480 480 --cacheBatchSize 12 --resume ./runs/forestlpr/checkpoints/model_best.pth.tar --level 1 6 11 --dataset intra --subset 5-03 --world_thresh 3 --time_thresh 100Anymal intra-evaluation
python -W ignore eval.py --img_size 480 480 --cacheBatchSize 12 --resume ./runs/forestlpr/checkpoints/model_best.pth.tar --level 1 6 11 --dataset intra --subset 3-02 --world_thresh 3 --time_thresh 100Here we provide the Code of Baselines for testing Wild-Places, helping others reproduce the results to compare.
Download the original code (or use git submodule) and the added xx_ script folder. The key files are inter-sequence.py and intra-sequence.py . Put the pretrained model into the checkpoints folder. Move xx_ script folder. (e.g., move transloc_scripts into TransLoc3D, and rename it to scripts)
TransLoc3D on Wild-Places, intra
Note: the cfg.model_cfg.quantization_size has been changed compared with reproduction in Wild-Places, for better performance.
cd scriptsig ../configs/transloc3d_baseline_cfg.py --databases $_PATH_TO_DATASET/Wild-Places/testing/K-04.pickle --run_names Karawatha/K-04 --save_dir ../dataset/Wild-Places/testing/ --database_featu
python intra-sequence.py --confres None --root $_PATH_TO_DATASET/Wild-Places --quan_size 1
quan_size is used to set cfg.model_cfg.quantization_size
Logg3D-Net on Botanic, intra
cd scripts
python intra-sequence.py --databases $_PATH_TO_DATASET/botanic/testing/5-03.pickle --run_names 1005-03/02 --save_dir ../dataset/botanic/testing/ --database_features None --root $_PATH_TO_DATASET/botanic --time_thresh 100 --world_thresh 3
Logg3D-Net on Wild-Places, inter
cd scripts
python inter-sequence.py --databases $_PATH_TO_DATASET/Wild-Places/testing/Venman_evaluation_database.pickle --queries $_PATH_TO_DATASET/Wild-Places/testing/Venman_evaluation_query.pickle --location_names Venman --query_features None --database_features None --root $_PATH_TO_DATASET/Wild-Places
MinkLoc3Dv2 on ANYmal, intra
use v2 configuration
cd scripts
python intra-sequence.py --databases $_PATH_TO_DATASET/anymal/testing/3-02.pickle --run_names anymal/03-2 --save_dir ../dataset/Wild-Places/testing/ --database_fe
atures None --root $_PATH_TO_DATASET/anymal --config ../config/config_baseline_v2.txt --model_config ../models/minkloc3dv2.txt --weights ../checkpoints/MinkLoc3Dv2.pth --world_thresh 3 --time_thresh 100
Scan-Context
cd scripts/eval/scan_context & python sc_intra_sequence.py --config ./sc_eval_config_anymal.yaml
MapClosure
python mc_intra_sequence.py --config mc_eval_config_wp.yaml
BevPlace
python eval_bp.py --dataset intra --cacheBatchSize 2 --img_size 480 480 --features_dim 8192 --subset K-03 --resume ./runs/bevplace/checkpoint_wp.tar --world_thresh 1
| Type | V-03 | V-04 | K-03 | K-04 | ANYmal | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| R@1 | F1 | R@1 | F1 | R@1 | F1 | R@1 | F1 | R@1 | F1 | ||
| ForestLPR | 76.53 | 64.15 | 82.33 | 78.62 | 74.89 | 65.01 | 76.73 | 81.97 | 71.87 | 81.45 | |
| Type | Inter-V | Inter-K | Validation | Botanic05-3 | Botanic05-6 | ||
|---|---|---|---|---|---|---|---|
| R@1 | R@1 | R@1 | R@1 | F1 | R@1 | F1 | |
| ForestLPR | 77.14 | 79.02 | 80.03 | 84.81 | 78.21 | 82.00 | 78.82 |
The retrieval video on ANYmal dataset.
demo.mp4
If you find the repo useful, please consider citing our paper:
@inproceedings{forestlpr,
title={ForestLPR: LiDAR Place Recognition in Forests Attentioning Multiple BEV Density Images},
author={Shen, Yanqing and Tuna, Turcan and Hutter, Marco and Cadena, Cesar and Zheng, Nanning},
booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
year={2025}
}
We would like to sincerely thank Wild-Places for the awesome released code.