SeqOT

The code for our paper accepted by IEEE Transactions on Industrial Electronics:

SeqOT: A Spatial-Temporal Transformer Network for Place Recognition Using Sequential LiDAR Data.

[IEEE Xplore TIE 2022] [arXiv]

SeqOT is a sequence-enhanced LiDAR-based place recognition method based on our previous work OverlapTransformer (OT) for RAL/IROS 2022.

Developed by Junyi Ma and Xieyuanli Chen.

Fig. 1 System overview with Haomo dataset and visualized evaluation on NCLT dataset. It can be seen that SeqOT is robust to driving directions.

News!

[2022-12] Our paper is accepted by IEEE Transactions on Industrial Electronics (TIE)!

Publication
Dependencies
How to Use
Data Preparation
License

Publication

If you use the code in your work, please cite our paper:

@ARTICLE{ma2022tie,
  author={Ma, Junyi and Chen, Xieyuanli and Xu, Jingyi and Xiong, Guangming},
  journal={IEEE Transactions on Industrial Electronics}, 
  title={SeqOT: A Spatial-Temporal Transformer Network for Place Recognition Using Sequential LiDAR Data}, 
  year={2022},
  doi={10.1109/TIE.2022.3229385}}

Dependencies

We use pytorch-gpu for neural networks.

An nvidia GPU is needed for faster retrival.

To use a GPU, first you need to install the nvidia driver and CUDA.

CUDA Installation guide: link
We use CUDA 11.3 in our work. Other versions of CUDA are also supported but you should choose the corresponding torch version in the following Torch dependences.

System dependencies:

sudo apt-get update 
sudo apt-get install -y python3-pip python3-tk
sudo -H pip3 install --upgrade pip

Torch dependences:
Following this link, you can download Torch dependences by pip:

pip3 install torch==1.10.2+cu113 torchvision==0.11.3+cu113 torchaudio==0.10.2+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

or by conda:

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

Other Python dependencies (may also work with different versions than mentioned in the requirements file):
```
sudo -H pip3 install -r requirements.txt
```

How to Use

We provide a training and test tutorial for NCLT sequences in this repository.

Training

You can start the two-step training with

cd train
python training_seqot.py 
python gen_sub_descriptors.py
python training_gem.py

Note that we only train our model using the oldest sequence of NCLT dataset (2012-01-08), to prove that our model works well for long time spans even if seeing limited data.

Test

The sub-descriptors of the database and query sequence have already been generated by python gen_sub_descriptors.py above. Here you can test SeqOT by

cd test
python test_gem_prepare.py
python test_seqot.py

Before training and test, please modify the params in the config.yml

data_root

range_image_database_root: path of reference range images in database sequence.
range_image_query_root: path of query range images in query sequence.

training_seqot

traindata_file: index for training. (link)
poses_file: poses of the trajectory of the database. (link)
height: height of range images.
width: width of range images.
seqlen: sequence length to generate sub-descriptors.
lr: learning rate for training feature extracter.
resume: whether resuming from the pretrained model below.
weights: pretrained feature extracter. (link)

gen_sub_descriptors

seqlen: sequence length to generate sub-descriptors.
weights: pretrained feature extracter. (link)

training_gem

traindata_file: index for training (same as the one of training_seqot).
poses_file: poses of the trajectory of the database (same as the one of training_seqot).
descriptors_file: sub-descriptors of database for training (generated by gen_sub_descriptors.py).
seqlen: sequence length to generate final-descriptors.
lr: learning rate for training the pooling module.
resume: whether resuming from the pretrained model below.
weights: pretrained pooling module. (link)

test_gem_prepare

sub_descriptors_database_file: sub-descriptors of database scans (generated by gen_sub_descriptors.py).
sub_descriptors_query_file: sub-descriptors of query scans (generated by gen_sub_descriptors.py).
seqlen: sequence length to generate final-descriptors (same as the one of training_gem).
weights: pretrained pooling module. (link)

test_seqot

groud_truth_file: ground truth containing correct loops. (link)

viz

poses_database: poses of the trajectory of the database. (link)
poses_query: poses of the trajectory of the query sequence. (link)

gen_training_index

poses_database: poses of the trajectory of the database. (link)
poses_query: poses of the trajectory of the query sequence. (link)
scan_database_root: path of reference .bin of the database. (link)

Data Preparation

Please use the following commands to prepare data. We have uploaded all the necessary files mentioned above except the range images which you can generate easily following our instruction.

Range Image Generation

Please use this script to generate range images of NCLT dataset. You need to modify the params including scan_folder and dst_folder.

Training Index Generation

Please use this script to generate training indices of NCLT dataset. You need to modify the params including poses_database, poses_query, and scan_database_root.

cd data_prepararion
python ./gen_training_index.py

Groud Truth Generation

Please use this script to generate groud truth loops of NCLT dataset.

cd data_prepararion
python ./gen_ground_truth.py

License

This project is free software made available under the MIT License. For more details see the LICENSE file.

BIT-MJY/SeqOT