Development of a PyTorch-Based Framework for Semantic Segmentation on Point Clouds

A short description of the proposal.

Requirements

  • M.Sc. in Machine Learning, Data Science, Computer Science
  • Strong knowledge of Python
  • Strong software development skills
  • Good concepts of data science, concerning data analysis, processing and machine learning

Description

The use of machine learning models for processing point clouds has become increasingly important in applications such as 3D object recognition, autonomous driving, and geospatial analysis. One of the key challenges in this domain is semantic segmentation, where each point in the cloud is assigned a specific class label. However, existing implementations often lack modularity and flexibility, making it difficult to adapt and extend them for different use cases. This internship aims to develop a new framework for training and inference of machine learning models specialized in semantic segmentation on point clouds, leveraging PyTorch and PyTorch Lightning to enhance efficiency, scalability, and ease of use. The primary goal of this internship is to refactor and merge two existing repositories widely adopted for 3D machine learning ( Open3D-ML and Pointcept ), replacing its current technology stack with PyTorch and PyTorch Lightning. The work will focus on improving code structure, modularizing key components, develop a detailed documentation and ensuring compatibility with state-of-the-art deep learning practices for semantic segmentation tasks. The resulting framework should be easy to extend, allowing researchers and developers to integrate new models and methodologies seamlessly. The internship will begin with an in-depth analysis of the existing repository, identifying its core functionalities, dependencies, and limitations. A structured refactoring process will then be carried out, transitioning the codebase to PyTorch while maintaining or improving its performance. PyTorch Lightning will be used to handle model training and inference, providing a cleaner and more efficient implementation.

Key aspects of the development process will include:

  • Designing a modular architecture that supports different model architectures for semantic segmentation.

  • Implementing efficient data loading and preprocessing techniques optimized for point cloud structures.

  • Integrating support for distributed training and model checkpointing using PyTorch Lightning.

  • Ensuring compatibility with widely used point cloud datasets and evaluation metrics specific to semantic segmentation.

By the end of the internship, a fully functional and well-documented framework for semantic segmentation on point clouds using PyTorch and PyTorch Lightning will be developed. This framework will serve as a foundation for future research and development, making it easier to experiment with new architectures and methodologies. Furthermore, this work could lead to a thesis project focusing on the development of an innovative machine learning approach for point cloud semantic segmentation, seamlessly integrated within the framework. This internship provides an opportunity to work at the intersection of deep learning and 3D data processing, contributing to the advancement of machine learning techniques for point cloud semantic segmentation. By refactoring an existing repository and adopting modern deep learning tools, this project aims to create a robust and scalable framework that can be widely used for research and practical applications.

Contacts