3D-Clustering
Description
The challenge of HAhRD project is to implement new algorithms to classify objects from 3D images-like coming from the data acquisition of the future sub-detector of CMS. This detector that contains about 6 million channels will be used to reconstruct the 3D clusters from hundreds of impinging particles arising from the proton-proton collisions within the Large Hadron Collider. The initial implementation based on an image processing algorithm is already exploited. We want in this proposal implement several Deep Neural Networks (DNN) architectures (in particular Convolutional Neural Networks - CNN) to classify clusters of points and develop a chain suite to analyze the classification performed by the DNN/CNN.
Task ideas and expected results
- Start with the original source code and data
- Implement one of the selected CNN architectures (GPUs can be used on our platform)
- Implement a validation tool chain to give results of the classification quality.
- Train the CNN with different objectives (different kind of objects to identify).
- Optimize the efficiency of the whole process.
- Extend the application to different CNN architectures or propose a software architecture
Requirements
Good C++/C skills, good python skills, familiar with GPUs if possible and visualization tools. Knowledge on machine learning or image processing or statistics would be appreciated.
Mentors
- Florian Beaudette, senior physicist
- Artur Lobanov, physicist
- Arnaud Chiron, computer scientist, CMS software expert
- Andrea Sartirana, computer scientist, Software environment expert (containers, …)
- Gilles Grasseau, computer scientist and image processing