The goal of this project is to apply, adapt, and further develop the recently proposed graph convolutional neural networks [1] to a practical problem: detecting objects in 3D point clouds. That problem arises in a number of practical scenarios, e.g. for autonomous cars to identify pedestrians from a LIDAR scan, or for governments to survey large areas for wildlife from aerial imagery taken by drones.

The project involves on one side to understand and further develop theoretical concepts recently introduced in deep learning: convolutional networks for ill-structured data, e.g. data lying on graphs. Such algorithms have been used to model and extract information from e.g. traffic on a road network or activations on a brain connectome. On the other side the project deals with real data and a concrete application: the detection of objects in 3D point clouds derived from aerial imagery. In essence, the student will have to master theoretical concepts and use them for a practical problem.

The student will i) learn about signal processing on graphs and its use to adapt deep learning to unstructured data, ii) apply the algorithm to an already studied problem to familiarize himself with the various intricacies of training deep nets, iii) play with the real point clouds provided by Picterra to understand the data and get a sense of what to expect, iv) construct a graph from the point clouds, v) apply the method to the problem of object recognition, and vi) tweak and improve the algorithm or pre-processing to attain maximal performance.

The ideal student knows about deep learning, is interested in the problem, knows how to engineer software, and is autonomous. During the project you will learn about one of the latest advances in deep learning and you will work at the edge of AI. Being a joint project between an academic laboratory — the LTS2 is involved in signal processing and machine learning on and with graphs — and a startup — Picterra is a company specialized in geo-information extraction from aerial and satellite imagery — , you will see both how research is conducted and how a company operates. Working desks are available at both locations. As the project is conducted in collaboration with Picterra, a startup based in the EPFL Garage, it can count as an internship.

References

[1] Defferrard, Michaël, Xavier Bresson, and Pierre Vandergheynst. "Convolutional neural networks on graphs with fast localized spectral filtering." Advances in Neural Information Processing Systems. 2016.

Metadata

  • Type: Master project
  • Supervisors: Michaël Defferrard from EPFL LTS2 and Frank de Morsier from Picterra
  • Professor: Pierre Vandergheynst from EPFL LTS2
  • Keywords: machine learning, deep learning, point clouds
  • Prerequisite: signal processing & machine learning background, python programming