[ICRA 2024] LPS-Net: Lightweight Parameter-shared Network for Point Cloud-based Place Recognition

Chengxin Liu, Guiyou Chen, Ran Song

Shandong University

Introduction

We propose a lightweight parameter-shared network (LPS-Net), which includes multiple bidirectional perception units (BPUs) to extract multi-scale long-range contextual information and parameter-shared NetVLADs (PS-VLADs) to aggregate descriptors. In the BPU, we design a parameter-shared convolution module (SharedConv) that can not only significantly compress the model but also enhance its ability to capture informative features. In the PS-VLAD, we replace half of the parameters used in the original NetVLAD with a trainable scalar and theoretically prove their equivalence. We designed variants of LPS-Net for three application scenarios. For limited computational resources, LPS-Net-S only includes one BPU. To balance between model performance and size, LPS-Net-M includes two BPUs. For scenarios requiring exceptional performance, LPS-Net-L includes three BPUs for higher accuracy. The experimental results demonstrate that our LPS-Net achieves the state-of-the-art on the point cloud based place recognition task while maintaining a highly lightweight model size.

Pre-requisites

• Python 3.8
• CUDA 11.1
• Pytorch 1.8.0
• Numpy 1.19.5
• Pandas 1.2.2
• Scikit-learn 1.2.1
• PyYAML 6.0
• Thop 0.1.1.post2209072238

Benchmark Datasets

• Download the zip file of the benchmark datasets found here and extract the folder to LPS_Net/benchmark_datasets.

• Generate pickle files. Note that the files only need to be generated once.

    cd dataldad/
  
    # For network training
    python generate_training_tuples_baseline.py
  
    # For network evaluation
    python generate_test_sets.py
  

Training and Evaluation

• Build the ops

  cd libs/pointops && python setup.py install && cd ../../
  

• Train the network

  # To train the LPS-Net-S
  python train.py --config configs/LPS_Net_S.yaml
  
  # To train the LPS-Net-M
  python train.py --config configs/LPS_Net_M.yaml
  
  # To train the LPS-Net-L
  python train.py --config configs/LPS_Net_L.yaml
  

• Evaluate the pre-trained network

  # To evaluate the LPS-Net-S
  python eval.py --config configs/LPS_Net_S.yaml
  
  # To evaluate the LPS-Net-M
  python eval.py --config configs/LPS_Net_M.yaml
  
  # To evaluate the LPS-Net-L
  python eval.py --config configs/LPS_Net_L.yaml
  

• Calculate the size of the model

  # To calculate the size of the LPS-Net-S
  python calculate_model.py --config configs/LPS_Net_S.yaml
  
  # To calculate the size of the LPS-Net-M
  python calculate_model.py --config configs/LPS_Net_M.yaml
  
  # To calculate the size of the LPS-Net-L
  python calculate_model.py --config configs/LPS_Net_L.yaml
  

Results

average recall@1%

Method	Oxford	U.S.	R.A.	B.D
PointNetVLAD	80.9	72.7	60.8	65.3
LPD-Net	94.9	96.0	90.4	89.1
PCAN	83.9	79.1	71.2	66.8
PPT-Net	98.1	97.5	93.3	90.0
MinkLoc3D	97.9	95.0	91.2	88.5
SVT-Net	97.8	96.5	92.7	90.7
EPC-Net	94.7	96.5	88.6	84.9
EPC-Net-L-D	92.2	87.2	80.0	75.5
LPS-Net-S (Ours)	96.4	97.0	92.3	89.1
LPS-Net-M (Ours)	97.3	98.6	94.4	92.4
LPS-Net-L (Ours)	97.6	99.1	95.5	92.3

Acknowledgement

Our code refers to PointNetVLAD and PPT-Net.