How does PolarNet compare to RangeNet++ for LiDAR segmentation?

PolarNet achieves higher mIoU (57.2% vs 52.2% on SemanticKITTI) and avoids KNN graphs for faster inference, but RangeNet++ might be easier to implement for some due to its different architecture.

How to train PolarNet on a custom LiDAR dataset?

Format your data to match SemanticKITTI's structure, adjust grid size parameters in training scripts, and modify data loaders; the README provides limited guidance, so expect coding effort for custom adaptations.

What hardware is needed to run PolarNet in real-time?

For training, a GPU with at least 12GB memory is recommended; inference can run on lower-end hardware, but performance depends on grid size and point cloud density.

Can PolarNet handle dynamic objects in autonomous driving?

Yes, it's designed for online segmentation of single scans, making it suitable for dynamic environments, but accuracy may vary based on dataset annotations and scene complexity.

Is there a pre-trained model for nuScenes available?

The README includes training code for nuScenes but doesn't specify pre-trained models; you'll likely need to train from scratch or fine-tune on that dataset.

GitHub repository

BSD-3-ClausePython

A lightweight neural network for near-real-time semantic segmentation of LiDAR point clouds using polar coordinate quantization.

GitHub

420 stars85 forks0 contributors

What is GitHub repository?

PolarNet is a neural network implementation for semantic segmentation of LiDAR point clouds, designed for online processing of single scans. It solves the problem of slow inference in existing methods by using a polar coordinate grid representation, avoiding graph construction and 3D convolutions. This enables near-real-time segmentation critical for applications like autonomous vehicles.

Target Audience

Researchers and engineers working on autonomous driving, robotics, or computer vision who need efficient LiDAR point cloud segmentation. It's particularly suited for those developing real-time perception systems.

Value Proposition

Developers choose PolarNet for its balance of high accuracy and fast inference, outperforming alternatives like SqueezeSegV2 and RangeNet++ on benchmark datasets. Its lightweight design avoids computational bottlenecks while maintaining state-of-the-art results.

Overview

Implementation for PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation (CVPR 2020)

Use Cases

Best For

Real-time LiDAR scan segmentation for autonomous vehicles

Related Projects

Community-curated · Updated weekly · 100% open source

Found a gem we're missing?

Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.

Submit a project Star on GitHub

Semantic segmentation of point clouds in robotics applications

Benchmarking new LiDAR segmentation methods against state-of-the-art

Processing large-scale outdoor LiDAR datasets like SemanticKITTI

Deploying efficient perception systems with limited computational resources

Research on grid-based representations for 3D point clouds

Not Ideal For

Offline batch processing of historical LiDAR datasets
Tasks requiring detailed 3D convolutions for volumetric feature extraction
Teams seeking plug-and-play deployment with minimal setup
Applications with non-LiDAR point clouds or indoor environments

Pros & Cons

Pros

Fast Inference Speed

Avoids KNN graph construction and 3D convolutions, enabling near-real-time processing essential for autonomous driving, as stated in the README's focus on online segmentation.

High Benchmark Accuracy

Achieves state-of-the-art mIoU scores on SemanticKITTI (57.2%), A2D2 (23.9%), and Paris-Lille-3D (53.3%), outperforming alternatives like SqueezeSegV2 and RangeNet++.

Efficient Grid Representation

Uses polar coordinate quantization to project points into a 2D grid, reducing computational complexity while maintaining segmentation quality, as described in the overview.

Multi-Dataset Support

Compatible with SemanticKITTI, nuScenes, and Paris-Lille-3D datasets, facilitating cross-dataset experimentation, with dedicated training scripts provided.

Cons

Complex Setup Process

Requires manual installation of dependencies like torch-scatter and dropblock, along with specific dataset folder structures, making initial deployment time-consuming.

High GPU Memory Demands

Training requires a TITAN Xp with 12GB memory; smaller GPUs may cause out-of-memory errors, necessitating grid size adjustments as noted in the training section.

Limited Generalization Beyond LiDAR

Tailored for outdoor LiDAR scans from specific datasets; adapting to other point cloud types or sensors would require significant code modifications.

Frequently Asked Questions

Home

LIDAR

GitHub repository

Open3D: A Modern Library for 3D Data Processing

Draco is a library for compressing and decompressing 3D geometric meshes and point clouds. It is intended to improve the storage and transmission of 3D graphics.

Stars7,334

Forks1,051

Last commit20 days ago

mmdetection3d

OpenMMLab's next-generation platform for general 3D object detection.

Stars6,441

Forks1,772

Last commit1 year ago

PCDet

OpenPCDet Toolbox for LiDAR-based 3D Object Detection.

Stars5,610

Forks1,453

Last commit8 months ago

#semantic-segmentation