This repository contains the code implementation for the paper "P-GAT : Pose-Graph Attentional Network for Lidar Place Recognition" accepted for publication in IEEE Robotics and Automation Letters (RA-L, 2024).

Utilising an attentional graph neural network, P-GAT compares (key)nodes between sequential and non-sequential sub-graphs for place recognition tasks as opposed to a common frame-to-frame retrieval problem formulation currently implemented in SOTA place recognition methods.

  author={Ramezani, Milad and Wang, Liang and Knights, Joshua and Li, Zhibin and Pounds, Pauline and Moghadam, Peyman},
  journal={IEEE Robotics and Automation Letters}, 
  title={Pose-Graph Attentional Graph Neural Network for Lidar Place Recognition}, 
  year={2024},
  volume={9},
  number={2},
  pages={1182-1189},
  doi={10.1109/LRA.2023.3341766}}

The arXiv version of P-GAT (Extended version of the RA-L paper) can be found here (link).

The code was tested using Python 3.9.4 with PyTorch 1.11.0. To set up the complete environment, we recommend using the following commands:

conda create -n `<myenv>` python=3.9.4
conda activate `<myenv>`

Replace the <myenv> by any name you want.

pip install -r requirements.txt
conda activate <myenv>

For the training and evaluation, P-GAT uses three datasets:

• Oxford RobotCar dataset
• In-house dataset
• MulRan dataset

We follow MinkLoc3D (link) to generate the descriptors for point clouds and save them in pickles. Oxford dataset can be downloaded from the PoinNetVLAD repository (link). MulRan dataset can be downloaded from (link). As an example, pickles generated for Oxford (based on MinkLoc3D) can be downloaded from (link) (Note: Dataset preparation for Oxford may require up to one day. For conveninece, the elements needed for evaluation on Oxford can be downloaded here (link))

We use datasets/dataset_generator_oxford.py to process the Oxford and in-house dataset for training and testing, datasets/dataset_generator_mulran.py to process DCC and riverside sub-datasets in MulRan dataset, and datasets/dataset_generator_kaist.py to process KAIST sub-dataset in MulRan.

A global graph and sub-graphs are generated inside the scripts. Note: The global graph and its corresponding adjacency matrix are exclusively used for training purposes. The usage of the scripts is:

usage: dataset_generator.py [-h] [--config_file FILE] ...

Generate dataset in graph structure

positional arguments:
  opts                Modify config options using the command-line

optional arguments:
  -h, --help          show this help message and exit
  --config_file FILE  path to config file

The arguments required by the script is maintained by the YAML configuration files. An example of the config file is configs/data_generator_config.yml, and an example of usage command is:

python dataset_generator_oxford.py --config <path-to-config>/data_generator_config.yml

We use training/train.py to train the model.

The default configuration is in the attentional_graph/config_init/defaults.py. The customised configuration is applied via configuration files in format of YAML. An example of config file in YAML can be found in configs/training_config.yml.

An example of training command is:

python training/train.py --config <path-to-config>/training_config.yml

Checkpoints listed below are the P-GAT models trained based on the descriptors generated by PointNetVLAD, MinkLoc3D and PPTNet on the Oxford and MulRan datasets.

We use training/test.py to evaluate the model.

Same as the training, the default configuration is in the attentional_graph/config/defaults.py, and the customised configuration is applied via configuration files in format of YAML. The example of config file in YAML can be found in configs/test_config.yml.

An example of testing command is:

cp <pre-trained-model.pt> .
python training/test.py --config <path-to-config>/test_config.yml

P-GAT has been developed based on the publically available lidar place recongition approaches including MinkLoc3D (link), PointNet-VLAD (link) and PPT-Net(link).

P-GAT is inspired by the attentional graph neural network developed in SuperGlue (link).