This repository is an open-source implementation of the ICRA 2022 paper: LoGG3D-Net: Locally Guided Global Descriptor Learning for 3D Place Recognition that won 2nd place at the General Place Recognition: City-scale UGV 3D Localization Competition - Round 1.
- Paper pre-print: https://arxiv.org/abs/2109.08336
- ICRA 2022 presentation: https://www.youtube.com/watch?v=HB6C6RHeYUU
We introduce a local consistency loss that can be used in an end-to-end global descriptor learning setting to enforce consistency of the local embeddings extracted from point clouds of the same location. We demonstrate how enforcing this property in the local features contributes towards better performance of the global descriptor in 3D place recognition. We formulate our approach in an end-to-end trainable architecture called LoGG3D-Net.
- [2023-05] Added new general dataloaders to help with training on custom datasets (such as Wild-Places).
- [2023-03] LoGG3D-Net has been extended for 6-DoF metric localization in SpectralGV RA-L 2023.
- [2023-03] Added support for PyTorch 2.0.
- [2022-05] Won 2nd place at the ICRA 2022 General Place Recognition Competition organized by AirLab, Carnegie Mellon University. Watch the invited talk here.
- [2022-05] Training code released.
- [2022-02] Evaluation code and pretrained models released.
- Joint constraints for local and global descriptors during training.
- Inference on high-resolution point-clouds using Sparse Point-Voxel convolution to capture fine-grained detail.
- Feature aggregation using higher-oder pooling to better capture the distribution of local features.
This project has been tested on a system with Ubuntu 22.04. Main dependencies include:
- Pytorch >= 1.13
- TorchSparse = 1.4.0
- Open3d >= 0.13.0
Set up the requirments as follows:
- Create conda environment with python:
conda create -n logg3d_env python=3.9.4
conda activate logg3d_env- Install PyTorch with suitable cudatoolkit version. See here:
pip3 install torch torchvision torchaudio
# Make sure the pytorch cuda version matches your output of 'nvcc --version'pip install -r requirements.txt- Install torchsparse-1.4.0
sudo apt-get install libsparsehash-dev
pip install --upgrade git+https://github.com/mit-han-lab/[email protected]- Install mpi4py:
conda install mpi4py
conda install openmpi- Download our pre-trained models from DropBox or cloudstor. Contains 7 checkpoints (6 for Kitti and 1 for MulRan) totalling 741.4 MB. Extract the content into
./checkpoints/:
wget -O checkpoints.zip https://cloudstor.aarnet.edu.au/plus/s/G9z6VzR72TRm09S/download
unzip checkpoints.zip- Download the KITTI odometry dataset, the MulRan dataset and set the paths in
config/eval_config.py. - For the KITTI odometry dataset, we are using the refined ground-truth poses provided in SemanticKITTI.
- For the MulRan dataset, create
scan_poses.csvfiles for each sequence using:
python ./utils/data_utils/mulran_save_scan_poses.pyBefore training:
- Do offline mining of positive pairs for both datasets (at 3m and 20m):
python utils/data_utils/kitti_tuple_mining.py
python utils/data_utils/mulran_tuple_mining.py- Set the number of GPUs available:
_NGPU=1Training:
- eg. Default training parameters on Kitti:
torchpack dist-run -np ${_NGPU} python training/train.py \
--train_pipeline 'LOGG3D' \
--dataset 'KittiPointSparseTupleDataset'- eg. Default training parameters on MulRan:
torchpack dist-run -np ${_NGPU} python training/train.py \
--train_pipeline 'LOGG3D' \
--dataset 'MulRanPointSparseTupleDataset'- See
config/train_config.pyfor all other training parameters.
For KITTI (eg. sequence 06):
python evaluation/evaluate.py \
--eval_dataset 'KittiDataset' \
--kitti_eval_seq 6 \
--checkpoint_name '/kitti_10cm_loo/2021-09-14_06-43-47_3n24h_Kitti_v10_q29_10s6_262450.pth' \
--skip_time 30For MulRan (eg. sequence DCC03):
python evaluation/evaluate.py \
--eval_dataset 'MulRanDataset' \
--mulran_eval_seq 'DCC/DCC_03' \
--checkpoint_name '/mulran_10cm/2021-09-14_08-59-00_3n24h_MulRan_v10_q29_4s_263039.pth' \
--skip_time 90Qualitative visualizations of top-1 retrievals on KITTI 08 and MulRan DCC 03:
Visualization of t-SNE embeddings of the local features extracted using our pre-trained model (on the CMU AirLab ALITA dataset).
If you find this work usefull in your research, please consider citing:
@inproceedings{vid2022logg3d,
title={LoGG3D-Net: Locally Guided Global Descriptor Learning for 3D Place Recognition},
author={Vidanapathirana, Kavisha and Ramezani, Milad and Moghadam, Peyman and Sridharan, Sridha and Fookes, Clinton},
booktitle={2022 International Conference on Robotics and Automation (ICRA)},
pages={2215--2221},
year={2022}
}
Functions from 3rd party have been acknowledged at the respective function definitions or readme files. This project was mainly inspired by the following: FCGF and SPVNAS.
For questions/feedback,
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent