3D Scene Completion with PointNet Autoencoder

This project started as an exploration of how neural networks can complete missing parts of 3D scenes — like reconstructing a broken vase or restoring incomplete scans. It’s built around a simple PointNet-style autoencoder trained on partial point clouds from ShapeNet.

🧠 What It Does

• Takes in partial 3D point clouds (like damaged or occluded objects).
• Learns to generate complete versions using a lightweight neural model.
• Evaluates performance using Chamfer Distance.
• Lets you visualize inputs and reconstructions to see how well it’s doing.

✨ Why I Built It

I’ve always found it fascinating how models can "fill in the blanks," whether it’s with text, images, or 3D data. This project gave me the perfect opportunity to explore that idea further, especially in the context of geometric deep learning. It allowed me to dive into working with autoencoders, data augmentation, and understanding 3D shapes. On top of that, it connects really well with the current research trends around generative models and 3D perception.

📂 Project Structure

• train.py: Main training loop.
• test.py: Evaluate the model on unseen data.
• inference.py: Run the model on a new .ply file.
• model.py: The PointNet-style autoencoder.
• dataset.py: Loads and occludes ShapeNet data.
• utils.py: Chamfer distance, visualizations, etc.
• config.py: Centralized configuration.
• visualize_results.py: Compare before/after shapes.
• requirements.txt: Dependencies.
• README.md: You’re reading it :).

📦 Dataset

Uses partial point clouds derived from ShapeNet. You’ll need to prepare or download .ply files and place them under ./data/ShapeNet/.

🧪 Training

To train the model, simply run:

python train.py

## 📊 Evaluation
To evaluate the model's performance, run:

```bash
python test.py

📊 Evaluation

To evaluate the model's performance, run:

python test.py

This will output the average Chamfer Distance, a simple way to measure shape similarity.

🧍 Inference & Try It Out

For inference on a new .ply file, run:

python inference.py --input_path path/to/your/file.ply --output_path path/to/save/reconstructed.ply

It will reconstruct the complete shape, and you can visualize the results using Open3D. If you want to visualize the results:

python visualize_results.py --original path/to/your/file.ply --reconstructed path/to/save/reconstructed.ply

This script will display the original and reconstructed 3D shapes for comparison.

🚀 Try It Out

Make sure you have the dependencies:

pip install -r requirements.txt

okay guyz you are ready to goo !

License

This project is licensed under the MIT License.