This repository contains the code for 3D Scene Reconstruction using Point Cloud Data, which predicts structured commands to recreate 3D scenes from point cloud inputs.
The model developed in this project uses neural networks to convert point cloud data into structured language commands (e.g., make_wall, make_door, make_window) for architectural environments. The model architecture is inspired by Meta's SceneScript, and it has been trained on the Aria Synthetic Environments (ASE) dataset, which was also provided by Meta.
Figure 1: Model architecture used for 3D scene reconstruction, inspired by Meta's SceneScript.
The ASE dataset consists of procedurally generated scenes with point cloud data, making it ideal for testing and training models for architectural and indoor environment reconstruction.
- Clone the repository:
git clone https://github.com/MariaSaleem6571/3d_SceneScript.git
- Navigate to the project directory:
cd 3d_SceneScript - Create a virtual environment:
conda create -n 3dscenescript python=3.12 conda activate 3dscenescript
- Install the required dependencies:
pip install -r requirements.txt
The training of the model can be done using the experiments.py script, where the hyperparameters can be configured in the experiment_config.py file.
To start training, simply run:
python experiments.pyFor testing the model, use the testing.py script.
To run the testing, execute the following command:
python testing.py