PointNet++ for Radar Point-Cloud Clutter Detection

A PyTorch implementation attempt of PointNet and PointNet++-style architectures for clutter detection on RadarScenes radar point-cloud data.

Project Status

This is a personal implementation and learning project exploring PointNet and PointNet++ ideas for radar point-cloud clutter detection. The code is not presented as a complete production-ready perception stack. Some scripts are experimental and may depend on local dataset paths.

Motivation

Radar point clouds are sparse, noisy, and irregular compared with camera images or LiDAR scans. This makes clutter detection a useful problem for autonomous driving and perception research.

PointNet and PointNet++ are relevant because they process unordered point sets directly and can learn spatial features without converting the point cloud into a dense grid.

Technical Overview

This project provides a PyTorch-based implementation featuring:

• PointNet-style feature extraction
• PointNet++-style set abstraction
• MLP blocks
• Optional T-Net style transformation module
• Dataset creation and preprocessing for RadarScenes
• Training and testing scripts

Architecture

RadarScenes Data
      |
      v
Dataset Creation / Sampling
      |
      v
Point Cloud Tensor
      |
      v
PointNet / PointNet++ Model
      |
      v
Per-point or scene-level feature extraction
      |
      v
Clutter Detection Output

PointNet++-style logic:

Input Points
      |
      v
Sampling / Grouping
      |
      v
Set Abstraction
      |
      v
Shared MLP
      |
      v
Feature Aggregation
      |
      v
Classifier / Prediction Head

Repository Structure

src/
  models/
    MLP.py
    Tnet.py
    pointnet.py
    pointnet_plus.py
    SetAbstraction.py
  data/
    dataset_creation.py
scripts/
  train_file.py
  test.py
docs/
  architecture.md
README.md
requirements.txt
.gitignore

Main Components

Component	File	Purpose
MLP block	src/models/MLP.py	Shared multi-layer perceptron utilities used in feature extraction
T-Net	src/models/Tnet.py	Optional learned transformation module inspired by PointNet
PointNet	src/models/pointnet.py	Baseline PointNet-style model
PointNet++	src/models/pointnet_plus.py	Hierarchical point-cloud model using set abstraction
Set Abstraction	src/models/SetAbstraction.py	Sampling/grouping and local feature aggregation logic
Dataset creation	src/data/dataset_creation.py	Dataset preparation utilities for RadarScenes
Training script	scripts/train_file.py	Training entry point
Test script	scripts/test.py	Testing / debugging entry point

Dataset

This project is intended for RadarScenes-style radar point-cloud data. Dataset files are not included in the repository. Local paths may need to be configured manually. Raw data should not be committed, and large generated arrays/checkpoints should be ignored.

Setup

git clone https://github.com/koukakios/Pointnet_plus_plus_implementation.git
cd Pointnet_plus_plus_implementation

python -m venv .venv
source .venv/bin/activate   # macOS/Linux
# or
.venv\Scripts\activate      # Windows

pip install -r requirements.txt

Usage

Basic usage examples:

python scripts/train_file.py
python scripts/test.py

Depending on the local dataset layout, paths inside the scripts may need to be updated before running.

Roadmap

• Implement PointNet-style baseline
• Implement PointNet++-style model components
• Add set abstraction module
• Add training/testing scripts
• Clean dataset path configuration
• Add command-line arguments for training settings
• Add reproducible experiment configuration
• Add evaluation metrics and saved result summaries
• Add unit tests for sampling/grouping operations
• Add documentation for RadarScenes preprocessing
• Add model checkpointing and experiment logging

Skills Demonstrated

• PyTorch
• Deep learning for point clouds
• PointNet / PointNet++ architecture
• Radar perception
• Autonomous-driving data processing
• Sampling and grouping for unordered point sets
• Model training pipeline design
• Python project organization

Limitations

• Experimental implementation
• Dataset not included
• No guaranteed reproducibility yet
• Some paths may be local-machine specific
• Not production-ready
• Not a complete autonomous driving perception stack

Portfolio Note

This repository is maintained as a portfolio and learning project showing practical work on point-cloud deep learning, radar perception, and PointNet++-style model implementation.