PointNetPGAP: PointNetPGAP-SLC: A 3D LiDAR-based Place Recognition Approach with Segment-level Consistency Training for Mobile Robots in Horticulture

Authors: T. Barros, L. Garrote, P. Conde, M.J. Coombes, C. Liu, C. Premebida, U.J. Nunes

RA-L Paper: https://ieeexplore.ieee.org/document/10706020

Changelog

Setember 2024: Published in IEEE Robotics and Automation Letters.

June 2024: Added training and testing conditions for the six sequences.

May 2024: First results on 4 sequences of HORTO-3DLM.

Installation

You can install PointNetGAP locally in your machine. We provide an complete installation guide for conda.

1. Create conda environment with python: conda create -n pr_env python=3.9.4

2. Activate conda environment conda activate pr_env

3. Install cuda 11.7.0 conda install -c "nvidia/label/cuda-11.7.0" cuda-toolkit

4. Install Pytorch 2.0.1 conda install pytorch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 pytorch-cuda=11.7 -c pytorch -c nvidia

5. Testing installation ....

6. Install sparse pip install --upgrade git+https://github.com/mit-han-lab/torchsparse.git@v1.4.0

6.1 sudo apt-get install g++

6.2 sudo apt-get install libsparsehash-dev

7. Install other dependencies pip install -r requirements.txt

Train

Default training script

python script_train.py

Costum training/testing

python train_knn.py  
        --network PointNetPGAPLoss # networl_name
        --train 1 # [1,0] [Train,Test] Train or test
        --dataset_root path/to/dataset/root # path to Dataset 
        --val_set 
        --memory RAM # [DISK, RAM] 
        --device cuda # Device
        --save_predictions path/to/dir # To save the predictions
        --epochs 200
        --max_points 10000
        --experiment experiment_name
        --feat_dim 16
        --eval_batch_size 15
        --mini_batch_size 1000
        --loss_alpha 0.5

Testing on Generated Descriptors

Default training script:

In the script_eval.py, edit:

dataset_root = path/to/dataset/root

resume_root = path/to/descriptors

Then, Run:

python script_eval.py

Download Descriptors and Predictions here

HORTO-3DLM Dataset

The HORTO-3DLM dataset comprises four sequences OJ22, OJ23, ON22, and SJ23; Three sequences from orchards, namely from apples and cherries; and one sequence from strawberries;

For t

3D Maps

Download HORTO-3DLM here

Citation:

@ARTICLE{10706020,
        author={Barros, T. and Garrote, L. and Conde, P. and Coombes, M.J. and Liu, C. and Premebida, C. and Nunes, U.J.},
        journal={IEEE Robotics and Automation Letters}, 
        title={PointNetPGAP-SLC: A 3D LiDAR-Based Place Recognition Approach With Segment-Level Consistency Training for Mobile Robots in Horticulture}, 
        year={2024},
        pages={1-8},
        doi={10.1109/LRA.2024.3475044}
  }