RPVNet

Introdution

This is an Non-official implement of RPVNet:A Deep and Efficient Range-Point-Voxel Fusion Network for LiDAR Point Cloud Segmentation. It is a fusion network based on three kinds of representation including Range image，Point Clouds and Voxels. For each Range or Voxel branch, it consist of a stem, four upblocks and four downblock. And point branch only uses four simple mlps bringing great efficiency and ability to extract fine-grained geometric features. What's more, Gated Fusion Module(GFM) was designed to adaptively measurse the importance of feature for each branch.

More details for this job for searching this paper.

Chinese comments Version: https://github.com/GuoPingPan/RPVNet/tree/chinese

Dependencies

• python 3
• pyyaml
• argparse
• torch
• torchvision
• tqdm
• numpy
• torchsparse

torchsparse can be install by flowing this

Data Preparation

SemanticKITTI

Please follow the instructions from here to download the SemanticKITTI dataset (both KITTI Odometry dataset and SemanticKITTI labels) and extract all the files in the sequences folder to /dataset/semantic-kitti. You shall see 22 folders 00, 01, …, 21; each with subfolders named velodyne and labels.

dataset
 - sequences
  - 00
    - velodyne
      - 000000.bin
      - 000001.bin
      - ...
    - labels
      - 000000.label
      - 000001.label
      - ...
      
    - poses.txt
      

Quick Start

1. ensure the dependencies
2. train

python train.py -d dataset/sequences  --log <path to save model> [--ckpt <path to pretrained model> --freeze_layers --device <cpu or cuda>]

3. test

python inference.py -d dataset/sequences  -ckpt <path to model> [--device <cpu or cuda>]

Reference

• paper: RPVNet: A Deep and Efficient Range-Point-Voxel Fusion Network for LiDAR Point Cloud Segmentation
• spvnas: https://github.com/mit-han-lab/spvnas
• SalsaNext: https://github.com/TiagoCortinhal/SalsaNext
• kprnet: https://github.com/DeyvidKochanov-TomTom/kprnet