Skip to content

README.md

Latest commit

 

History

History
120 lines (79 loc) · 2.88 KB

README.md

File metadata and controls

120 lines (79 loc) · 2.88 KB

Point Cloud Semantic Segmentation (PandaSet + PointNet)

This repository provides a minimal, clean, and extendable ML pipeline for 3D point cloud semantic segmentation using the PandaSet LiDAR dataset. It implements a complete workflow from dataset acquisition → preprocessing → model architecture → training stub.

The goal is to provide a working baseline that can later be upgraded to advanced 3D sparse convolution architectures (e.g., MinkowskiNet / SparseConv U-Net).

🚀 Features

  • Full dataset download via Python (Kaggle API)
  • Custom PandaSet loader (point clouds + semantic labels)
  • PointNet segmentation baseline implemented in PyTorch
  • Minimal training loop (forward → loss → backward)
  • Configurable number of sequences (use subsets for fast prototyping)
  • Fully reproducible and dependency-light

📦 Setup

pip install -r requirements.txt

Create a .env file in the project root:

KAGGLE_USERNAME=your_username
KAGGLE_KEY=your_key

🗂 Dataset Download (PandaSet)

Run:

python download_pandaset.py

This script:

  • Authenticates using your .env
  • Downloads the full PandaSet dataset (~33GB) from Kaggle
  • Unzips everything under ./pandaset/

PandaSet is released under CC-BY-NC-SA, and Kaggle enforces attribution automatically via the LICENSE file.

Model Architecture — PointNet Baseline

The baseline model is a PointNet segmentation network implemented from scratch:

  • Input: (N, 4) points → (x, y, z, intensity)
  • MLP + shared Conv1D layers
  • Global feature aggregation
  • Per-point classification head

This provides a simple and reliable foundation before migrating to:

  • SparseConv U-Net
  • MinkowskiNet
  • KPConv
  • PointNet++
  • etc.

🏋️ Training

python pointnet_model.py

This runs:

  • Dataset loading
  • Minimal data analysis
  • One epoch of training (stub)
  • Prints batch loss

Designed for rapid validation and code completeness rather than full training.

📊 Expected Baseline Performance

  • PointNet on outdoor LiDAR: ~35–50% mIoU
  • SparseConv U-Net (production target): ~60–75% mIoU

PointNet is intentionally chosen for simplicity and compatibility with the assessment constraints.

🧱 Project Structure

point_cloud_segmentation/
│
├── download_pandaset.py        # Kaggle dataset downloader
├── pointnet_model.py           # Dataset loader + PointNet + training stub
├── MODEL_CHOICES.md            # Architecture justification & notes
├── requirements.txt            # Dependencies
└── .gitignore

📘 License Notice

PandaSet is provided under the CC-BY-NC-SA 4.0 license. Any commercial use must follow Scale.ai's licensing terms.

This repository contains no dataset files — users must download them via Kaggle using the provided script.