1.  Projects for spring 2010

1.1  Music Similarity Measures to Navigate through the Montreux Jazz Festival Library

The Montreux Jazz Festival library is the largest testimony of live music recorded on the same stage, both in audio and video, for the past 43 years (more than 4'000 bands among others were filmed in Montreux resulting in 10'000 recording tapes). EPFL and Montreux sounds have created a unique and one of a kind, high resolution digital archive of the Montreux Jazz Festivals, with EPFL as exclusive licensee for scientific research and educational use. The goal of this project is to exploit this huge archive using signal processing techniques in order to find and discover interesting music titles by deploying content-base retrieval strategies. In recent years , there has been a surge of interest in the Area of Music Information Retrieval (MIR) which success depends largely on the existence of relevant music similarity descriptors. Using these music descriptors, it is possible to efficiently select or pick-up titles from a large database of music titles, which are relatively similar (in sense of perceptual criteria like the tempo, the rhythm, the timbre, etc) to other related music tracks. The student will have to first study the relevance of these descriptors applied on the Montreux Jazz festival database, according to some objective and subjective criteria which are yet to be defined. The second part of this project consists in using these similarity measures so as to build a graph of the database titles, and to develop some algorithms to navigate through this graph of titles.

Project type: Semester (1 or 2 students intended for) / Mater Thesis project
Project Status: FREE
Requirements: Matlab programming, notions of signal/audio processing and graph theory
Supervisors: Dr Simon Arberet simon.arberet@epfl.ch (Contact person), Prof. Pierre Vandergheynst

1.2  Efficient Non-Local relationship estimation

Following the success of the Non-Local (NL) Means denoising algorithm (a very simple yet highly efficient image restoration algorithm), NL methods have become extensively studied and used for various image processing tasks. They suffer however from one major drawback : their conceptual simplicity leads, when implemented naively, to poorly efficient algorithms. In short, the simple straightforward implementation of an NL method leads to long, very long, computation times.

Over the recent years, different approaches have been proposed to tackle this difficulty, for example by treating the problem of NL relationship as a Nearest-Neighbour search endorsed by a data-driven structure (a kd-tree) or a feature space-driven structure (partitioning structures similar to quadtrees). The goal of this project is to compare the performance of these 2 approaches, along with other methods from the Machine Learning community, for various Non-Local problems. It will lead to the definition of guidelines on how to choose the correct method for a given problem, and possibly to the definition of an universal fast NL relationship estimation method.

Project type: Semester / Diploma project
Project Status: FREE
Requirements: C / Matlab programming, interest for algorithmic and design, basic notions of data/signal processing
Supervisors: Prof. Pierre Vandergheynst
Assistant: Emmanuel d'Angelo (Contact person)

1.3  Accelerated acquisition for magnetic resonance imaging.

Magnetic resonance imaging (MRI) is a well known non-invasive medical imaging technique to visualize the internal structure of the body. In many MRI applications, accelerating the acquisition process is of great interest. To achieve this goal, one may either develop methods to acquire faster all the data or reduce the amount of them, while preserving the required reconstruction quality.

On the one hand, compressed sensing is a recent theory which aims at merging data acquisition and compression. It shows that it is possible to recover sparse or compressible signals very accurately from a number of linear measurements far smaller than the one required by the Nyquist-Shannon theorem. The reconstruction quality depends on some properties of the measurement matrix. In very recent works, we showed that this reconstruction quality can be drastically enhanced by a "spread spectrum technique" consisting in the introduction of a pre-modulation in the acquisition scheme. On the other hand, "parallel imaging" uses multiple receiver coils and the data for each coil are acquired in parallel. The number of measurements may be reduced because each coil image are weighted differently. SMASH (Sodickson et al.), SENSE (Pruessmann et al.) and SPIR-iT (Lustig et al.) are some famous algorithms to reconstruct the image in this context.

The goal of the project resides in comparing the performance and analyzing the complementarity of the spread spectrum technique and parallel imaging.

Project type: Diploma project
Student: Mr. Mohammad Hamed Izadi
Requirements: Matlab/C programming, good notions of signal/image processing
Supervisors: Dr Yves Wiaux, Prof. Pierre Vandergheynst
Assistant: Gilles Puy (Contact person)

1.4  Shear estimation in galaxy images affected by gravitational lensing

In the context of the concordance cosmological model, our universe is filled in with 25% of dark matter and 74% of dark energy, while visible matter only represents 1% of the energy content. Unveiling the nature dark matter and dark energy represents a central issue for cosmology today. Weak gravitational lensing is the process in which the light from distant galaxies is bent by the gravity of mass located along the line of sight, hence causing a shape distortion (shear) of these galaxies. Statistical analysis of the shear provides essential information on the dark matter distribution in different epochs of the universe and on the properties of dark energy.

The shear estimation from observed galaxy images is an arduous task. Indeed, the original shape of galaxies is poorly known and observation introduces a convolution kernel, pixelization effects, and noise. The aim of this project is to develop a powerful statistical method, possibly in a Bayesian framework, to solve this very challenging inverse problem. A large set of simulations was made available for the related GRavitational lEnsing Accuracy Testing 2008 (GREAT08) Challenge (http://www.great08challenge.info/). The illustrative picture below is borrowed from the official GREAT08 document:

This work is proposed in the context of new collaborations of the signal processing laboratory at the international level.

Project type: Diploma project
Project Status: FREE
Targeted students: the project is open to interested students both in physics and engineering
Requirements: Matlab/C programming, good notions of signal/image processing, additional interest for statistical analysis and cosmology
Supervisors: Dr Yves Wiaux, Prof. Pierre Vandergheynst

1.5  Depth estimation from a single plenoptic picture

The Signal Processing Laboratory (LTS2) and Microelectronic Systems Laboratory (LSM) are developping a new generation of imager called Panoptic Camera. Emulating the biological visual system of certain insects, this device is composed of a large number of independent conventional imagers layered in spherical geometry (right figure) so as to capture a full field of view (omnidirectional) image. Unlike other omnidirectional cameras though, the panoptic camera is also a polydioptric system: each imager has a proper distinct focal point. This characteristic allows to capture information about the scene simultaneously from different imagers with overlapping field of view.

In this project we will make use of the unique capabilities of the panoptic system to interpolate light rays measured on a single shot and generate several virtual omnidirectional images of the scene from slightly different view points. This ability is in deep connection with modern concepts like light fields rendering, plenoptic vision and epipolar geometry. The resulting omnidirectional images will then be used to estimate the depth of the scene in all directions around the camera. The work will be carried on synthetic and real images acquired by a first generation prototype of the panoptic system.

Project type: Diploma project
Project Status: FREE
Requirements: Matlab programming, good notions of signal/image processing, the desire to work in a multidisciplinary team on disruptive technology
Supervisors: Prof. Pierre Vandergheynst
Assistant: Luigi Bagnato (Contact person)

1.6  Panoptic camera calibration

The second prototype of the Panoptic Camera (see previous project for a detailed description) is now ready. An important step in this phase of the project is an efficient way to model and estimate all the causes of non-ideal behavior of the device.

In this project we will find an efficient (semi-)automatic algorithm to calibrate the Panoptic camera using real data acquired by the device. The purpose of the calibration procedure is twofold:

1. Model and correct lens distortions for each camera.
2. Correct the errors due to the misposioning of the cameras with respect to their ideal positions on the hemisphere.

Project type: Semester Project/Diploma project
Project Status: FREE
Requirements: Matlab programming, notions of signal/image processing, the desire to work in a multidisciplinary team on disruptive technology
Supervisors: Prof. Pierre Vandergheynst
Assistant: Luigi Bagnato (Contact person)

2.  Projects for fall 2009

2.1  Tracking and Structure from Motion from Omnidirectional Sequences

Robust tracking and structure from motion (SFM) are fundamental problems in computer vision that have important applications for robot visual navigation and other computer vision tasks. Structure from motion refers to the process of finding the three-dimensional structure of the scene from the analysis of a video sequence. In this project we want to integrate a robust tracker to an existing framework for depth estimation from consecutive frames of an omnidirectional video. Depending on the skills and the personal interests we could propose the development of software for 3D scene visualization.

Links: http://en.wikipedia.org/wiki/Structure_from_motion

Project type: Diploma project
Project Status: Assigned to Andreas Weishaupt
Requirements:: Notion of signal processing and computer vision. Programming languages: c++ and Matlab.

Assistant: Luigi Bagnato (Contact person)

3.  Selected former projects

3.1  Denoising of cosmic strings with wavelets - [G. Puy - Diploma project - Fall 2008]

This project relates to the definition of an efficient denoising algorithm for the detection of physical defects in images of the universe (see figure on the right).

In the context of the concordance cosmological model, the structures of the universe originate in primordial Gaussian adiabatic energy density perturbations. However, theories for the unification of the fundamental interactions strongly suggest that phase transitions occurred in the early universe, which produced topological defects. Cosmic strings are the line-like version of such defects, which would leave their imprint in the cosmic microwave background (CMB) radiation in terms of straight temperature steps. The corresponding non-Gaussian component of the CMB induced by strings would simply add to the standard Gaussian component. The present status of observations still allows a small fraction of structures in the universe to find their origin in topological defects.

The aim of this project is to enhance a recently proposed denoising method for the detection of cosmic strings in CMB data with wavelets. It will be performed in the context of ongoing international collaborations of the laboratory. From data acquisition to analysis, issues to be addressed may be summarized by the following keywords: compressed sensing of the cosmic string signal - denoising of the cosmic string signal on the plane or on the sphere with wavelets - evidence for cosmic strings in terms of a Bayesian or a frequentist statistical analysis of the denoised signal.

Requirements: Matlab/C programming, good notions of signal/image processing, additional interest for statistical analysis and cosmology
Supervisors: Dr Yves Wiaux, Prof. Pierre Vandergheynst