Enhancing 3D Point Cloud Classification

Overview

ModelNet‑R is our refined version of the ModelNet40 dataset. By resolving labeling inconsistencies, removing near‑2D data, and improving class differentiation, we provide a cleaner, more reliable benchmark for 3D point cloud classification.

Point‑SkipNet is a lightweight, graph‑based neural network designed to efficiently extract features from 3D point clouds using smart sampling, grouping, and skip connections. This combination achieves state‑of‑the‑art accuracy with lower computational overhead.

Challenges in ModelNet40

Inconsistent labeling and ambiguous class definitions
Presence of nearly 2D objects lacking volumetric depth
Size mismatches causing classification confusion

Our Solution: ModelNet‑R & Point‑SkipNet

ModelNet‑R Dataset: A refined dataset featuring corrected labels, removal of 2D artifacts, and enhanced class differentiation.
Point‑SkipNet Architecture: A lightweight neural network that leverages efficient point sampling, grouping, and skip connections for improved feature extraction.

Point‑SkipNet Architecture — Figure: The Point‑SkipNet architecture

Figure: Sample and Group Module for efficient point sampling

Dataset

Place these datasets in the data folder:

data/modelnet40_normal_resampled
data/modelnet40_ply_hdf5_2048

Note: The modelnet40_normal_resampled dataset will be converted to modelnetR_normal_resampled during processing. Please back up the original data if needed.

Training & Testing

To train Point‑SkipNet:

cd Point-SkipNet
dataset="modelnetR" bash run_train.sh

To test the model:

cd Point-SkipNet
dataset="modelnetR" bash run_test.sh

Experimental Results

Our experiments demonstrate significant improvements using ModelNet‑R and Point‑SkipNet:

ModelNet: Overall Accuracy ~92.29%, Mean Class Accuracy ~89.84%
ModelNet‑R: Overall Accuracy ~94.33%, Mean Class Accuracy ~92.93%

Resources

Read the Paper

Explore the full methodology and results in our publication.

View Paper

GitHub Repository

Access the complete code, dataset, and documentation.

Visit GitHub

YouTube Video

Watch our detailed presentation on YouTube.

Watch Video

Download Presentation

Get the PDF presentation for an in-depth overview.

Download PDF

Acknowledgements

This project builds upon the valuable contributions of the research community. Special thanks to:

PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space (arXiv)
GitHub Implementation: PointNet++ PyTorch
ModelNet Dataset: The original benchmark from 3D ShapeNets (PDF)