Immense amounts of high resolution data are now routinely produced thanks to recent advances in Electron Microscopy (EM) imaging. Our lab recently focused on developing automatic 3D segmentation techniques for such large datasets. Our algorithm is a graph-based approach where the graph nodes encode a predefined probability distribution. Segmenting a dataset amounts to assigning a label to each graph node so that the probability distribution is maximized. This process is also known as inference in statistics and probability theory.
A recent paper (Distributed Message Passing for Large Scale Graphical Models [link]) proposed a distributed algorithm for inference on large scale graphical models. It can handle large problems efficiently by distributing and parallelizing the computation and the memory requirements among many different machines. The goal of this project is to investigate the use of this algorithm on one of our large high resolution EM dataset. The source code is available on request so the student won't have to re-implement the full approach.
The final goal of this project is to produce 3D reconstruction like in the video above but for a very large EM dataset .
Students should have experience in C++ programming and machine learning.
30% Theory, 60% Implementation (C++), 10% Practical experiments
For further information, send an e-mail regarding the project.
Contacts: Aurélien Lucchi (office BC 306, tel 3 75 18),
Yunpeng Li (office BC 301, tel 3 81 95), and
Kevin Smith.