3DPointCloud

This Repository contains code for reconstructing 3D point clouds using the Occupancy Predictions of a small and sparse subsets of points. This work is based on Lionar, Stefan, et al. "Dynamic Plane Convolutional Occupancy Networks" Proceedings of the IEEE/CVF Winter Conference of Applications of Computer Vision, 2021.

Table of Contents

• Installation
• Architecture

Installation

1. Clone the repository:

   git clone https://github.com/EugenioBugli/3DPointCloud.git

2. Install dependencies:

   pip install -r <Folder>/3DPointCloud/requirements.txt

3. You can run the Code directly from the Notebook

Architecture

The Architecture used has an Encoder-Decoder structure :

Encoder

1. ResNetPointNet
2. Plane Predictor
3. UNet

Decoder

1. ResNet
2. Occupancy Predictor

The obtained occupancy predictions are then used to reconstruct the mesh by using the Multiresolution IsoSurface Extraction (MISE) and the Marching Cubes algorithm