Scattering Networks for Scene Labeling

Description

Scattering networks [1,2] extract characteristic features from signals by recursively applying the composition of the following three operations: convolution with a set of filter functions, a non-linearity, and a sub-sampling step. Such networks, combined with a classifier (such as, e.g., support vector machines) were successfully employed in a number of classification tasks [1,3]. In contrast to traditional convolutional neural networks which learn the filters from training data [4], scattering networks use pre-defined filters such as wavelets [1], curvelets [2], or shearlets [2]. While a scattering network-based classifier arguably leads to less flexible machine learning models than convolutional network-based classifiers which learn the filters, they potentially allow for faster implementations thanks to the structure of the pre-defined filters (e.g., tensorized wavelet filters).

Consequently, this project shall explore the application of scene labeling, which aims at assigning a class label such as “street”, “tree”, or “building” to every pixel of an image and is employed, e.g., in situational awareness systems [4]. Such systems often demand low-complexity and low-power scene labeling algorithms and the scattering transform may help to meet these needs.

The goal of this project is to develop a scattering network-based classifier for scene labeling and to realize a GPU implementation thereof (followed by VLSI/FPGA, if time permits).

Status: Available

Contact/Supervision:

• Michael Tschannen (michaelt ät nari.ee.ethz.ch)
• Thomas Wiatowski (withomas ät nari.ee.ethz.ch)
• Lukas Cavigelli
• Michael Lerjen (mlerjen ät nari.ee.ethz.ch)

Prerequisites

Matlab, C programming, linear algebra

Character

0%-20% Theory

80%-100% Programming

Professor

tbd ↑ top

References

↑ top