Pipeline

INSTALLATION

Set up environment

• python 3.9.4
• Cuda 11.7
• pytorch 2.0.1
• open3d
• torchpack

Create conda environment with python:

conda create -n pr_env python=3.9.4

conda activate pr_env

Install cuda 11.7.0

conda install -c "nvidia/label/cuda-11.7.0" cuda-toolkit

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

Install Troch Sparse

Note: Tested on Ubuntu 20.04 and 22.04

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

Install remaining libs

pip install -r requirements.txt

Download

Dataset

Download the dataset from https://github.com/Cybonic/HORTO-3DLM

direct link

Checkpoints

The checkpoints can be downloaded from SPVSoAP3D_iros24.zip.

SPVSoAP3D_iros24.zip contains additionally the descriptors and results obtained with SPVSoAP3D on all sequences. SPVSoAP3D_iros24.zip has the following strucutre:

24SPVSoAP3D_iros24.zip
├── READ-ME.md # file explaining the meaning of each file 
├── GT23 
├── OJ22
├── OJ23
├── ON22
├── ON23
└── SJ23
    ├── checkpoints.pth # trained weights
    ├── descriptors.torch # descriptors
    └── place
        ├── results_precision.csv # precision 
        ├── results_recall.csv # recall 
        ├── loops.csv # predicted loops
        ├── scores.csv # similarity scores  
        └── target.csv # ground-truth loops

For more details on the xxx.csv files, please read READ-ME.md file in the 24SPVSoAP3D_iros24.zip.

Testing

Run eval on all sequnces

1. Edit the script_eval.py file, adding the chackpoint and dataset paths

script_eval.py:

chkpt_root = 'path/to/checkpoints'

dataset_root = 'path/to/dataset'

# Select the option to run from checkpoints.pth or precomputed descriptors.torch 
resume  = "checkpoints.pth" # <-- choise [checkpoints.pth, descriptors.torch]

2. Run the file

python script_eval.py

Run eval on a single sequence

resume choice: [checkpoints.pth or descriptors.torch]

python eval_knn.py 
    --val_set <seq>
    --resume <path/to/24SPVSoAP3D_iros24>/<seq>/<resume> 
    --save_predictions  <path/to/save/predictions>

Note: We have noticed that performance changes when using GPU with different cuda versions.

Training

Training Triplets

• Anchor (Green)
• Positive (Blue)
• Negatives (Red)

ON23	GTJ23

Script to train all sequences

1. First edit script_train.py to set the path to the dataset and other parameters

2. Then, run script

python script_train.py 

Train a single sequence

python train_knn.py 
    --val_set <seq>
    --network <model>
    --resume <chkpt_root>/<model>/<seq>/<resume>
    --memory <memory>  # [DISK, RAM] 
    --device cuda' # [cuda,cpu]
    --save_predictions  <path/to/save/predictions> # Save predictions
    -e <experiment name>
    --roi 0 # point cloud cropping range [m] around the LiDAR origin 
    --augmentation 0  # [0,1] 
    --pcl_norm 0 # chcoice [0,1] Normalization of the point clouds so points are within [0,1]