Enhancing 3D Point Cloud Classification with ModelNet‑R and Point‑SkipNet

Overview

3D point cloud classification is pivotal in applications like autonomous driving, robotics, and augmented reality. However, existing benchmarks such as ModelNet40 suffer from inconsistent labeling, the presence of 2D data, size mismatches, and ambiguous class definitions.

Our contributions:

• ModelNet‑R Dataset: A refined version of ModelNet40 that resolves labeling inconsistencies, removes low-quality (almost 2D) data, adjusts size disparities, and sharpens class boundaries.
• Point‑SkipNet: A lightweight graph‑based neural network that leverages efficient sampling, neighborhood grouping, and skip connections to achieve state‑of‑the‑art classification accuracy with a reduced computational footprint.

For full details, please refer to the paper on arXiv.

---

Homepage

Official implementation of the paper "Enhancing 3D Point Cloud Classification with ModelNet-R and Point-SkipNet",

accepted as an Oral presentation at IPRIA 2025 🔥

The paper will be available soon...

---

Problem Statement: Limitations of ModelNet40

ModelNet40, while widely used, contains:

• Inconsistent labeling: Many samples are mislabeled or ambiguous.
• 2D artifacts: A significant portion of the data is nearly flat, lacking volumetric depth.
• Size mismatches: Normalization often causes objects of different real-world scales to appear similar.
• Poor class differentiation: Similar geometries between classes (e.g., flower pot vs. vase) lead to confusion.

---

Proposed Solution: ModelNet‑R & Point‑SkipNet

Introducing ModelNet‑R

By addressing the issues of ModelNet40, ModelNet‑R provides a cleaner and more reliable benchmark for 3D point cloud research. It improves both the training process and the evaluation of classification models.

Lightweight Classification with Point‑SkipNet

Point‑SkipNet is designed to:

• Efficiently sample and group points using a dedicated module.
• Leverage skip connections to preserve both global and local features.
• Reduce computational overhead while maintaining high classification accuracy.

Architecture Overview

The architecture is modular:

• Sample and Group Module: Uses farthest point sampling (FPS) and ball queries to extract meaningful local neighborhoods.
• Feature Aggregation: Combines local geometric features with skip connections for robust global representation.

---

Experimental Insights

Experiments demonstrate that refining the dataset (ModelNet‑R) coupled with the efficient design of Point‑SkipNet results in significant performance improvements. For instance:

• Point‑SkipNet on ModelNet:

  • Overall Accuracy (OA): ~92.29%
  • Mean Class Accuracy (mAcc): ~89.84%

• Point‑SkipNet on ModelNet‑R:

  • Overall Accuracy (OA): ~94.33%
  • Mean Class Accuracy (mAcc): ~92.93%

These results underscore the importance of both dataset quality and computational efficiency.

---

Prerequisites & Setup

Tested configurations include:

• Configuration 1:

  • OS: Ubuntu 22.04.3 LTS
  • GPU: NVIDIA GeForce RTX 2080 Ti
  • CUDA: 12.0
  • PyTorch: 2.1.1+cu121
  • Python: 3.10.12

• Configuration 2:

  • OS: Ubuntu 22.04.5 LTS
  • GPU: NVIDIA GeForce RTX 3080 Ti
  • CUDA: 12.2
  • PyTorch: 1.13.1+cu117
  • Python: 3.10.12

---

Data Preparation

Place the following 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 when you run the dataset creation script. It is recommended to back up the original data if needed.

---

Creating ModelNet‑R

To generate the refined dataset:

cd Creating_ModelNet-R
h5=false bash creat_modelnetR.sh  # Use this if you do not require the h5 format
h5=true bash creat_modelnetR.sh   # Use this if you require the h5 format

Ensure that the modelnet40_normal_resampled dataset is present in the data folder before running the script.

---

Training & Testing Point‑SkipNet

Training

Train Point‑SkipNet on your chosen dataset:

cd Point-SkipNet
dataset="modelnetR" bash run_train.sh   # Train on ModelNet‑R (refined dataset)
dataset="modelnet" bash run_train.sh    # Train on ModelNet (original dataset)

Testing

Test the trained model using the provided checkpoints:

cd Point-SkipNet
dataset="modelnetR" bash run_test.sh   # Test on ModelNet‑R
dataset="modelnet" bash run_test.sh    # Test on ModelNet

---

The results of the models trained on the ModelNet-R dataset

Model	OA	mAcc	Param(M)
PointNet	91.39	88.79	3.47
PointNet++ (ssg)	94.02	92.40	1.47
PointNet++ (msg)	94.06	91.80	1.74
Point-NN	84.75	77.65	0
DG-CNN	94.03	92.64	1.8
CurveNet	94.12	92.65	2.04
PointMLP	95.33	94.30	12.6
Point-SkipNet(proposed)	94.33	92.93	1.47

Citation

If you use this work, please cite:

@inproceedings{2025paper,
  author    = {Mohammad Saeid and Amir Salarpour and Pedram MohajerAnsari},
  title     = {Enhancing 3D Point Cloud Classification with ModelNet-R and Point-SkipNet},
  booktitle = {Proceedings of IPRIA 2025},
  year      = {2025},
}

---

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 by Qi et al.

• PointNet++ PyTorch Implementation
  Repository by yanx27

• ModelNet Dataset
  Originally presented in 3D ShapeNets: A Deep Representation for Volumetric Shapes

We also acknowledge all related works cited in the paper for inspiring this study.

---

Future Work

Future enhancements include:

• Refining all classes in ModelNet40 for uniform consistency.
• Incorporating techniques to retain size-related features lost during normalization.
• Extending validations of Point‑SkipNet to additional real‑world datasets such as ScanObjectNN and ShapeNet.

---

This repository offers an integrated approach to improve both dataset quality and model efficiency for 3D point cloud classification. Enjoy exploring the project, and feel free to reach out through the project homepage for any inquiries!