Reflectivity is All You Need!: Advancing LiDAR Semantic Segmentation

[Paper]

ICRA25_2305_VI_i.1.mp4

Summary

This repository explores the benefits of incorporating calibrated intensity (reflectivity) in learning-based LiDAR semantic segmentation frameworks. By leveraging reflectivity alongside raw intensity measurements, our model exhibits improved performance, particularly in off-road scenarios.

rxyzi represents model trained on raw intensity data. rxyzn represents model trained on reflectivity data.

Generating reflectivity data

Generate reflectivity data for Rellis-3D:

python utils/data_generator.py

Modify the dataset and output file path.

Data generators for Semantic-Kitti and Semantic-POSS can be found in /utils.

Modified Rellis-3D dataset used for training and testing. Download.

(a) Illustrates spherical projection of LiDAR with raw intensity as pixel values. (b) Calibrated for range and angle of incidence. (c) Calibrated for range, angle of incidence and near-range effect.

SalsaNext

Original Salsanext and modified versions config files can be found in:

cd ./train/tasks/semantic/config/arch

The modified versions of salsanext mentioned in paper is:

salsanext_rxyzi.yml
salsanext_rxyzirn.yml
salsanext_rxyzn.yml
early_ga_detach.yml (*learning reflectivity*)

Evaluate SalsaNext

Download Pretrained models.Google Drive.

Make sure you have installed all the python dependencies.

pip install -r ./requirements.txt

Inorder to evaluate Salsanext:

bash run_salsanext_eval.sh 

Edit the paths for dataset and pretrained models in evaluate.sh

To generate LiDAR images for SAM Model

python utils/lidar_image_generator.py

The LiDAR specifications and paths needs to be modified in the code.

The GUI used in the video for SAM based LiDAR annotations can be found here

There are several helpers and utilities in /utils.

Related Research

Citation

@article{VISWANATH202543,
title = {Reflectivity Is All You Need!: Advancing LiDAR Semantic Segmentation},
journal = {IFAC-PapersOnLine},
volume = {59},
number = {3},
pages = {43-48},
year = {2025},
note = {12th IFAC Symposium on Intelligent Autonomous Vehicles IAV 2025},
issn = {2405-8963},
doi = {https://doi.org/10.1016/j.ifacol.2025.07.008},
url = {https://www.sciencedirect.com/science/article/pii/S2405896325003477},
author = {Kasi Viswanath and Peng Jiang and Srikanth Saripalli}
}