PointNet 3D Classification (ModelNet40)

PointNet is a classic neural network for learning directly from unordered 3D point clouds by applying shared MLPs to each point and using a symmetric aggregation (global max pooling) to obtain a permutation-invariant global feature. This project implements PointNet from scratch in PyTorch and trains it on ModelNet40 using torch_geometric.datasets.ModelNet.

Project Structure

pointnet-3d-classification/
├── data/           # data loading utilities
├── models/         # PointNet architecture
├── utils/          # helper functions
├── train.py        # training script
├── evaluate.py     # evaluation script
├── requirements.txt
└── README.md

Architecture (Text / ASCII Diagram)

          Point Cloud Input
            x: [N, 3]
                 |
          +---------------+
          |  Input T-Net  |  predicts A_in: [3, 3]
          +---------------+
                 |
         x = A_in * x
                 |
          +---------------+
          | Shared MLP #1 |  Conv1d(1x1): 3 -> 64
          +---------------+
                 |
          +---------------+
          | Feature T-Net |  predicts A_feat: [64, 64]
          +---------------+
                 |
         x = A_feat * x
                 |
          +---------------+
          | Shared MLP #2 |  Conv1d(1x1): 64 -> 128 -> 1024
          +---------------+
                 |
        Global Max Pool over N points
                 |
               [1024]
                 |
          FC(1024->512) -> Dropout
                 |
          FC(512->256)  -> Dropout
                 |
             FC(256->K)
                 |
           logits: [K=40]

How to Install

1) Python + virtual environment

Use Python 3.10+:

python -m venv .venv
source .venv/bin/activate

2) Install PyTorch

Install a PyTorch build appropriate for your machine (CPU or CUDA) from the official site.

3) Install PyTorch Geometric (torch-geometric)

torch-geometric has binary dependencies. Follow the official installation guide for your PyTorch version: https://pytorch-geometric.readthedocs.io/

4) Install project dependencies

pip install -r requirements.txt

Dataset Download + Preprocessing

The code automatically downloads ModelNet40 via torch_geometric.datasets.ModelNet. ModelNet is stored as meshes; we preprocess it into a fixed-size point cloud using: torch_geometric.transforms.SamplePoints(num=<num_points>)

Train

python train.py \
  --data-root data/modelnet \
  --num-points 1024 \
  --epochs 50 \
  --lr 0.001 \
  --batch-size 32 \
  --output-dir checkpoints/pointnet-modelnet40

During training, the script logs per-epoch training/validation loss and accuracy and saves the best model checkpoint to: checkpoints/pointnet-modelnet40/best_model.pth It also writes a CSV log to: checkpoints/pointnet-modelnet40/training_log.csv

Evaluate

python evaluate.py \
  --data-root data/modelnet \
  --num-points 1024 \
  --checkpoint checkpoints/pointnet-modelnet40/best_model.pth

The evaluation prints per-class accuracy and overall test accuracy.

Results (Actual)

These are the actual results from a quick CPU run with normalized point clouds.

Run settings

• Num points: 512
• Epochs: 8
• Train/val split: 0.9 / 0.1 (from the training set)
• Checkpoint: checkpoints/run_reasonable8/best_model.pth

Command used

python train.py --data-root data/modelnet --num-points 512 --epochs 8 --lr 0.001 --batch-size 16 --num-workers 0 --output-dir checkpoints/run_reasonable8 --val-split 0.1 --cpu
python evaluate.py --data-root data/modelnet --num-points 512 --checkpoint checkpoints/run_reasonable8/best_model.pth --batch-size 16 --num-workers 0 --cpu

Num Points	Best Val Accuracy	Overall Test Accuracy
512	84.35%	79.13%

Why PointNet Matters for 3D Understanding

3D point clouds have no inherent order, but many standard neural layers assume ordered inputs. PointNet solves this by using:

1. shared weights across points (so each point is encoded the same way), and
2. a symmetric pooling function (global max pooling) to aggregate point features into a permutation-invariant representation.

This makes PointNet a foundational architecture for tasks like classification and part segmentation in 3D.