Student Projects

LTS5 OPEN SEMESTER AND MASTER PROJECTS – Autumn 2019

SEMESTER PROJECT PROPOSALS

1. Facial Attributes Modeling

Human face has always been of particular interest in the computer graphics community. Because of its complexity, modeling lifelike synthetic objects is challenging. A variety of approaches have been proposed such as statistical models (i.e. principal component analysis models) or blendshapes models. Moreover, tackling the variation in terms of population (*i.e. identity*) and expression at the same time in a generic 3D model increases the difficulty.

With conventional modelling technics, the detailed facial attributes such as the wrinkles are lost in the process. Moreover these mid-frequencies informations are important for photo-realistic expressions generation. Therefore an explicit model can be jointly used to augment the original face model to recover them. Furthermore, photo-realistic rendering shading rely on depth cue given by ambient occlusion (AO). The ambient occlusions define how the light is attenuated at a specific location and is therefore dependent on the geometry of the face.

The goals of this project are in two folds. Build a model of the detailed facial attributes distribtion from high resolution meshes. Then explore the relation between the geometry and its corresponding ambient occlusions map in order to generate them it (i.e. regresse ambient occlusions from shape coefficients).

Requirements: The project will be implemented in Python / C++ so good knowledge is required. Previous experience in one or several of the following topics would be a plus: image processing, computer vision or machine learning.

Assistant: Christophe Ecabert (christophe.ecabert@epfl.ch)

Supervisor: Prof. Jean-Philippe Thiran

2. Tissue microstructure estimation using deep networks trained from dictionary-based methods

Diffusion-Weighted Magnetic Resonance Imaging (DW-MRI) captures the patterns of water displacement in the neuronal tissue and allows noninvasive investigation of the tissue microstructure. A number of biophysical models have been proposed to relate the tissue organization with the observed diffusion signals. One important microstructure property is the axon diameter distribution in the brain’s white matter. Such models use the diffusion signal with three compartments that are characterized by distinct diffusion properties. In this project, the student will use a deep network-based approach that performs end-to-end estimation of the axon diameters indexes based on a dictionary-based method developed by our group.

Requirements: Python.

Tools to use Keras and Theano, Convex optimization python libraries.

Depending on the student background the main activities involve:

  • Test and train Deep Networks.
  • Modify and test Network parameters.
  • Understand the basics of Diffusion-Weighted Magnetic Resonance Imaging
  • Test and improve

Supervisor: Prof. Jean-Philippe Thiran
Assistant: Jonathan Rafael-Patiño

3. 3D tissue segmentation and reconstruction.

Image segmentation is the process of assigning a label to every pixel in an image such that pixels with the same label share certain characteristics, e.g. tissue characteristics. When applied to a stack of images, typical in medical imaging, the resulting contours after image segmentation can be used to create 3D reconstructions with the help of interpolation algorithms. In this project, the student will use semi-automatic tools to train and classify image stacks of brain’s white matter tissue, to then use the classified labels to generate a 3d mesh model of the tissue compartments.

Requirements: Matlab or Python.

Tools to use: itk-snap segmentation tool and Blender.

Supervisor: Prof. Jean-Philippe Thiran
Assistant: Prof. Giorgio Innocenti, Jonathan Rafael-Patiño.

4. GPU-based ultrasound imaging on a smartphone

Among the diagnostic imaging technologies, ultrasound (US) imaging is the safest and least expensive one, which makes it one of the most commonly used in medical practice. Many current US devices are bulky machines not suitable for remote diagnosis. Recently, portable US devices have gained a significant interest, especially in point of care US such as emergency, non medicalized areas, telemedicine etc.

At LTS5, we are developing advanced methods for ultrafast US image reconstruction either based on iterative algorithms or more recently on deep neural network architectures. Due to the need for real-time imaging and thanks to their high parallelizability, some of these methods have already been implemented on high-end GPUs. Since we are targeting portable platforms, the next step is to make these techniques available on smartphones. To do so, two options may be investigated: iOS with its Metal 2 API, and Android with the Vulkan API.

The goal of this project is to develop a smartphone-based demonstrator of US imaging using the embedded GPU. With embedded platforms come power and complexity limitations that will have to be addressed. The student will have to:
1) Familiarize with ultrasound imaging and parallel computing
2) Study and choose the most appropriate smartphone platform and API
3) Implement a standard US image reconstruction algorithm
4) Study trade-offs between image quality and computation time/complexity

Depending on the evolution of the project other steps can be investigated such as:
– Interface the demonstrator with a real ultrasound scanner
– Implement advanced image reconstruction algorithms

Requirements: Good knowledge in C++ and parallel computing. Skills in iOS/Android and/or signal processing is a plus.

Supervisor: Prof. Jean-Philippe Thiran

Assistants: Florian Martinez (florian.martinez@epfl.ch) and Dimitris Perdios (dimitris.perdios@epfl.ch)

5. Dynamic image analysis for Cervical Cancer Detection: methods and Android demonstrator – analyse d’images dynamiques pour la détction du cancer du col de l’uterus: méthodes et démonstreteur Android

Cervical cancer is a major concern in public health, both in developed and developing countries. Especially in this later context, the availability of well-trained experts is limited, and computer-aided diagnosis is clearly needed for large-scale screening.

This project will focus on the analysis of dynamic image sequences (videos) of the cervix under a contrast agent: the visual inspection with acetic acid (VIA) is known as one of the reference methods to detect cervical cancer. However the human eye has limited capabilities in assessing the time evolution of the appearance of the cervix after administration of the contrast agent. Therefore, in this phase of the project, the intensity curves of each pixel of the video will be analyzed by a machine-learning algorithm. A synthetic image displaying the results will be produced as a help for the diagnosis. Moreover, as a simple and portable tool is needed, especially in developing countries, an Android application will be developed to perform this analysis and display the results on mobile platforms.

This project will be realized in close collaboration with the Department of Gynecology of the Geneva University Hospital.

Requirements: C++, image processing, skills in development for Android would be ideal.
Supervisor: Prof. Jean-Philippe Thiran

6. Apizoom – deep learning to quantify the Verroa parasite in honey bee hive images

Varroa mites are recognized as the biggest pest to honey bees worldwide, and are believed to be the single largest contributing factor in the modern-day decline of honey bees due to their ability to transmit diseases, resulting in death or severe deformity of the pupae.

Verroa on honey bees.

Detecting and quantifying the presence of Verroa in a beehive is therefore crucial to treat the infection appropriately and as early as possible, and image analysis appears very useful in this problem.

In this project, we propose to develop an image analysis to detect and count Verroa cadavers who felt down on a plate below the beehive, as a non-intrusive way to quantify the presence of the Verroa. High definition images will be capture and Deep Learning techniques will be investigated here, to design and train a Convolutional Neural Network (CNN) to detect the Verroa and distinguish it from other wastes. See the following video for more details on the project (in French).

Depending on the evolution of the project, several steps will be investigated:

  • Development of the Deep Learning method
  • Training on a collection of annotated images
  • Test and improve

And possibly:

  • Study integration in mobile phones
  • Study the development of a web-based analysis system.

This project is jointly proposed with the company Apizoom (Fribourg, Switzerland).

Responsible: Prof. J.-Ph. Thiran

7. CleanCityIndex – A Deep Learning based system to localize and classify wastes on the streets

A review of major European cities places “urban cleanliness” as a top priority for the authorities, as it directly impacts the concern and satisfaction of their citizens and the attractiveness of their economy and tourism. Littering quantification is an important step toward improving urban cleanliness. When human interpretation is too cumbersome or in some cases impossible, an objective index of cleanliness could reduce the littering by awareness actions.

The goal of this project is to propose a fully automated computer vision application for littering quantification, based on images taken from the streets and sidewalks. We employ a deep learning based framework to localize and classify different types of wastes.

In this project, the student will be involved into study and develop new detection algorithms and investigates deep learning based object tracking techniques for this specific case. The student is expected to be familiar with Python and Tensorflow.

Assistant: Saeed Rad (saeed.rad@epfl.ch)
Supervisor: Prof. Jean-Philippe Thiran

8. Computational pathology for automated analysis of histopathologic scans

Computational pathology is a state-of-the-art technology that aims to diagnose cancer and distinguish tissue components (e.g. nuclei, tumour) which has seen great improvements in recent years due to the advancement of convolutional neural networks (CNN) based diagnosis systems. However, automated analysis of histopathology whole-slide images is impeded by the scanner-dependent variance such as stain inconsistency introduced in the slide scanning process. In addition, CNNs are not the best suited for large scale (i.e. millions of pixels) multi-resolution histopathology whole slide images. Finding computationally efficient solutions to automatically analyze these images remains an open challenge. The goal of this project is to develop a computer-aided diagnosis (CAD) system, which can be used  for a number of applications where histopathologic images are captured from different scanners.

In this project, the student will be involved into study and develop CNN-based algorithms for histopathology-related applications including cancer detection and classification of the subtype from histopathologic scans . The student is expected to be familiar with Python and Tensorflow and/or Pytorch.

Assistant: Dr Behzad Bozorgtabar (behzad.bozorgtabar@epfl.ch)

Supervisor: Prof. Jean-Philippe Thiran

MASTER PROJECT PROPOSALS 

MEDICAL IMAGING PROJECTS

1. Tissue microstructure estimation using deep networks trained from dictionary-based methods

Diffusion-Weighted Magnetic Resonance Imaging (DW-MRI) captures the patterns of water displacement in the neuronal tissue and allows noninvasive investigation of the tissue microstructure. A number of biophysical models have been proposed to relate the tissue organization with the observed diffusion signals. One important microstructure property is the axon diameter distribution in the brain’s white matter. Such models use the diffusion signal with three compartments that are characterized by distinct diffusion properties. In this project, the student will use a deep network-based approach that performs end-to-end estimation of the axon diameters indexes based on a dictionary-based method developed by our group.

Requirements: Python.

Tools to use Keras and Theano, Convex optimization python libraries.

Depending on the student background the main activities involve:

  • Test and train Deep Networks.
  • Modify and test Network parameters.
  • Understand the basics of Diffusion-Weighted Magnetic Resonance Imaging
  • Test and improve

Supervisor: Prof. Jean-Philippe Thiran
Assistant: Jonathan Rafael-Patiño

2. 3D tissue segmentation and reconstruction.

Image segmentation is the process of assigning a label to every pixel in an image such that pixels with the same label share certain characteristics, e.g. tissue characteristics. When applied to a stack of images, typical in medical imaging, the resulting contours after image segmentation can be used to create 3D reconstructions with the help of interpolation algorithms. In this project, the student will use semi-automatic tools to train and classify image stacks of brain’s white matter tissue, to then use the classified labels to generate a 3d mesh model of the tissue compartments.

Requirements: Matlab or Python.

Tools to use: itk-snap segmentation tool and Blender.

Supervisor: Prof. Jean-Philippe Thiran
Assistant: Prof. Giorgio Innocenti, Jonathan Rafael-Patiño.

3. Advanced ultrafast Ultrasound image reconstruction – joint master project with the CREATIS Lab in Lyon

Ultrasound medical imaging (ultrasound) allows rapid and real-time imaging of certain parts of the human body. One of the major drawbacks of ultrasound imaging is the quality of the images obtained. Thanks to the expansion of computing resources, it is now possible to propose and use new models of the problem to consider the image obtained in a new light.

With this in mind, we want to take advantage of a direct model recently developed by the Créatis laboratory (University of Lyon) and the skills of the LTS5 laboratory (EPFL) in inverse problems to achieve ultra-fast ultrasound imaging with a high image quality. .

The project will consist first of all to take in hand the direct model for the emission in plane waves. This linear model will then be reversed in the case of a single plane wave, taking into account the properties of the medium, to obtain an ultrasound image. A practical implementation on a research ultrasound system will be carried out to validate the theoretical developments.

The project will be completed partly at EPFL-LTS5 and partly in Lyon. Practical modalities will be discussed with the interred candidates.

Supervisor: Jean-Philippe Thiran

4. Deep learning for ultrasound data-rate reduction

Ultrasound (US) is a widely used medical imaging modality mostly because of its non-invasive and real-time characteristics. Recent advances in US imaging (e.g. ultrafast imaging, 3D imaging, elastography, functional imaging etc.) gave rise to a crucial challenge: dealing with the huge amount of data that has to be transferred and processed in real-time. To address this problem, the LTS5 is focusing on two main aspects:
1) Maximizing the image quality for a given amount of data using advanced image reconstruction methods
2) Minimizing the data-rate to reach a given image quality using advanced sampling and compression strategies

This project proposes to address the second point of the list using the latest advances in deep learning for data compression. Deep neural networks (DNN) have recently drawn many interest in the context of image and video compression. The most successful architectures are trained as an encoder-decoder. Once trained, the data are firstly encoded into a low-dimensional space (i.e. latent space), in which some quantization might occur, before being transferred. The decoder is then capable of recovering data close to the original.

During this project, the student will investigate relevant network architectures and compare them to state-of-the-art data-rate reduction techniques, in the context of US signals. In particular the student will have to:
1) Study state-of-the-art data-rate-reduction approaches for US signals
2) Familiarize with deep learning approaches for data-rate reduction
3) Study, select and optimize relevant network architectures for US data-rate reduction
4) Optimize training procedures (e.g. generalization, loss function, etc.)
5) Target real-time applications (i.e. network complexity vs data-rate reduction trade-offs)
6) Test, validate and compare network architectures against state-of-the-art approaches on quantitative metrics

Requirements: Strong knowledge of signal processing, machine learning, convex optimization is a plus. Skills in Python (TensorFlow / PyTorch).

Supervisor: Prof. Jean-Philippe Thiran

Assistants: Dimitris Perdios (dimitris.perdios@epfl.ch) and Adrien Besson (adrien.besson@epfl.ch)

5. Deep learning for enhanced ultrasound image reconstruction

Ultrasound (US) is a widely used medical imaging modality mostly because of its non-invasive and real-time characteristics. Recent advances in US imaging (e.g. ultrafast imaging, 3D imaging, elastography, functional imaging etc.) gave rise to a crucial challenge: dealing with the huge amount of data that has to be transferred and processed in real-time. To tackle this challenge, the LTS5 is focusing on two main aspects:
1) Maximizing the image quality for a given amount of data using advanced image reconstruction methods
2) Minimizing the data-rate to reach a given image quality using advanced sampling and compression strategies

This project targets the first point of the list. In order to maximize the image quality, the image reconstruction process can be formulated as an ill-posed inverse problem. Many algorithms have been developed to solve such a problem, e.g. ADMM, LASSO, primal-dual forward-backward. All these algorithms are iterative and involve thresholding operations which depend on one or several hyperparameters that must be manually tuned for optimal reconstruction. Recently deep neural networks (DNN) have emerged as an alternative to the classical algorithms leading to impressive results. DNN have been used for different purposes. They can model the algorithms themselves leading to a network where each layer represents an iteration of the algorithm. DNN can also be used as a denoiser on the initial low-quality image.

During the project, the student will continue the investigation of deep learning methods as a way to overcome the drawbacks of regularized iterative methods. Current results in US image enhancement using DNN are very promising (see figure below) but many questions remain open. The student will have to:
1) Familiarize with deep learning approaches for solving inverse problems
2) Study and optimize relevant network architectures for this problem
3) Target real-time applications (i.e. study architecture complexity vs image quality trade-offs)
4) Study the impact of the dataset (e.g. simulated vs experimental, data augmentation, robustness, etc.)
5) Test, validate and compare network architectures on quantitative metrics

Requirements: Strong knowledge of signal processing, machine learning, convex optimization is a plus. Skills in Python (TensorFlow / PyTorch).

Supervisor: Prof. Jean-Philippe Thiran

Assistants: Dimitris Perdios (dimitris.perdios@epfl.ch) and Adrien Besson (adrien.besson@epfl.ch)

6. GPU-based ultrasound imaging on a smartphone

Among the diagnostic imaging technologies, ultrasound (US) imaging is the safest and least expensive one, which makes it one of the most commonly used in medical practice. Many current US devices are bulky machines not suitable for remote diagnosis. Recently, portable US devices have gained a significant interest, especially in point of care US such as emergency, non medicalized areas, telemedicine etc.

At LTS5, we are developing advanced methods for ultrafast US image reconstruction either based on iterative algorithms or more recently on deep neural network architectures. Due to the need for real-time imaging and thanks to their high parallelizability, some of these methods have already been implemented on high-end GPUs. Since we are targeting portable platforms, the next step is to make these techniques available on smartphones. To do so, two options may be investigated: iOS with its Metal 2 API, and Android with the Vulkan API.

The goal of this project is to develop a smartphone-based demonstrator of US imaging using the embedded GPU. With embedded platforms come power and complexity limitations that will have to be addressed. The student will have to:
1) Familiarize with ultrasound imaging and parallel computing
2) Study and choose the most appropriate smartphone platform and API
3) Implement a standard US image reconstruction algorithm
4) Study trade-offs between image quality and computation time/complexity

Depending on the evolution of the project other steps can be investigated such as:
– Interface the demonstrator with a real ultrasound scanner
– Implement advanced image reconstruction algorithms

Requirements: Good knowledge in C++ and parallel computing. Skills in iOS/Android and/or signal processing is a plus.

Supervisor: Prof. Jean-Philippe Thiran

Assistants: Florian Martinez (florian.martinez@epfl.ch) and Dimitris Perdios (dimitris.perdios@epfl.ch)

7. Dynamic image analysis for Cervical Cancer Detection: methods and Android demonstrator – analyse d’images dynamiques pour la détction du cancer du col de l’uterus: méthodes et démonstreteur Android

Cervical cancer is a major concern in public health, both in developed and developing countries. Especially in this later context, the availability of well-trained experts is limited, and computer-aided diagnosis is clearly needed for large-scale screening.

This project will focus on the analysis of dynamic image sequences (videos) of the cervix under a contrast agent: the visual inspection with acetic acid (VIA) is known as one of the reference methods to detect cervical cancer. However the human eye has limited capabilities in assessing the time evolution of the appearance of the cervix after administration of the contrast agent. Therefore, in this phase of the project, the intensity curves of each pixel of the video will be analyzed by a machine-learning algorithm. A synthetic image displaying the results will be produced as a help for the diagnosis. Moreover, as a simple and portable tool is needed, especially in developing countries, an Android application will be developed to perform this analysis and display the results on mobile platforms.

This project will be realized in close collaboration with the Department of Gynecology of the Geneva University Hospital.

Requirements: C++, image processing, skills in development for Android would be ideal.

Supervisor: Prof. Jean-Philippe Thiran

8. Master project in industry – Real time inverse planning in radiotherapy and radiosurgery by convex optimization and deep learning

Radiotherapy and radiosurgery are therapeutic approaches to treat cancerous tumors by focal X-ray irradiation, either through multiple low-dose sessions (radiotherapy) or a single high dose session (radiosurgery). In this technique, high dose conformity is achieved by focusing multiple irradiation beams onto the target. Planning a radiotherapy treatment consists thus in defining the position, incidence and intensity of the individual beams which can deliver sufficient radiation to a tumour while both sparing critical organs and minimizing the dose to healthy tissue. This is typically done with the help of so-called inverse planning software. However, the existing systems are slow and largely sub-optimal, since they involve complicated non-convex optimization techniques.

Intuitive Therapeutics SA, a start-up company located in St-Sulpice, next to EPFL, in collaboration with the LTS5, has developed a completely new approach for radiotherapy/radiosurgery inverse planning, relying on convex optimization and GPU acceleration, to achieve interactive real-time inverse planning. Although already very fast, the underlying optimization algorithm requires setting several important parameters to achieve optimal performances, which is a complex task.

In this project, we propose to develop the concept of learning those parameters by deep learning to achieve extremely fast convergence of the optimization algorithm. For confidentiality reasons we cannot disclose more details here, but feel free to contact us to know more about this project.

Supervisor: Prof. Jean-Philippe Thiran

9. Model-based optimal inversion of haemodynamic data – master project in industry at Leman Micro Devices (EPFL Innovation Parc)

Context: LMD’s V-Sensor and its software are designed to make a measurement of blood pressure by analysing the data collected over a period of typically 45 to 60 seconds. It works in a way that is equivalent to the cuff of a conventional automatic blood pressure meter:

  • the user varies the pressure applied to the fingertip from below diastolic pressure to above systolic pressure
  • the software extracts the optical PPG signal for each beat
  • the optical signals are mapped against the pressure applied to the finger
  • the software finds the applied pressures that are equal to diastolic and systolic pressures

This procedure therefore finds an average of diastolic and systolic pressures over the measurement time.

Project: A model-based inversion could transform the optical data to find the instantaneous blood pressure in the artery throughout the beat, using the full data set to optimise accuracy and efficiency and to enable new diagnostics:

  • From the instantaneous blood pressure, it is possible to measure the variation of the blood pressure between beats [1].
  • It is also possible to transform the data to find the blood pressure at other places in the body, including the aorta[2].

The aim of the project is to investigate these transformations and the accuracy of the results.

In practice the transformation will probably be more accurate for beats recorded with an applied pressure a little below diastolic. It might therefore be desirable to use a 2-stage data collection:

  • stage 1 as now to find diastolic and systolic pressures and the relationship between optical signal, applied pressure and arterial pressure
  • stage 2 to collect data at approximately constant applied pressure to find instantaneous blood pressure over several beats.

Technical: Data can be captured using LMD’s sensors and software and can be analysed using any preferred tools – Matlab or Python are the obvious candidates. It will require familiarity with signal processing, mathematical modelling and data processing, and also programming a user interface.

[1]    Hocht C “Blood Pressure Variability: Prognostic Value and Therapeutic Implications” ISRN Hypertension 2013, 398485 http://dx.doi.org/10.5402/2013/398485

[2]    Fetics et all “Parametric Model Derivation of Transfer Function for Noninvasive Estimation of Aortic Pressure by Radial Tonometry” IEEE Trans Biomed Eng, Vol. 46, No 6, Jun I999

12. Computational pathology for automated analysis of histopathologic scans

Computational pathology is a state-of-the-art technology that aims to diagnose cancer and distinguish tissue components (e.g. nuclei, tumour) which has seen great improvements in recent years due to the advancement of convolutional neural networks (CNN) based diagnosis systems. However, automated analysis of histopathology whole-slide images is impeded by the scanner-dependent variance such as stain inconsistency introduced in the slide scanning process. In addition, CNNs are not the best suited for large scale (i.e. millions of pixels) multi-resolution histopathology whole slide images. Finding computationally efficient solutions to automatically analyze these images remains an open challenge. The goal of this project is to develop a computer-aided diagnosis (CAD) system, which can be used  for a number of applications where histopathologic images are captured from different scanners.

In this project, the student will be involved into study and develop CNN-based algorithms for histopathology-related applications including cancer detection and classification of the subtype from histopathologic scans . The student is expected to be familiar with Python and Tensorflow and/or Pytorch.

Assistant: Dr Behzad Bozorgtabar (behzad.bozorgtabar@epfl.ch)

Supervisor: Prof. Jean-Philippe Thiran

COMPUTER VISION PROJECTS

​​​​10. Facial Attributes Modeling

Human face has always been of particular interest in the computer graphics community. Because of its complexity, modeling lifelike synthetic objects is challenging. A variety of approaches have been proposed such as statistical models (i.e. principal component analysis models) or blendshapes models. Moreover, tackling the variation in terms of population (*i.e. identity*) and expression at the same time in a generic 3D model increases the difficulty.

With conventional modelling technics, the detailed facial attributes such as the wrinkles are lost in the process. Moreover these mid-frequencies informations are important for photo-realistic expressions generation. Therefore an explicit model can be jointly used to augment the original face model to recover them. Furthermore, photo-realistic rendering shading rely on depth cue given by ambient occlusion (AO). The ambient occlusions define how the light is attenuated at a specific location and is therefore dependent on the geometry of the face.

The goals of this project are in two folds. Build a model of the detailed facial attributes distribtion from high resolution meshes. Then explore the relation between the geometry and its corresponding ambient occlusions map in order to generate them it (i.e. regresse ambient occlusions from shape coefficients).

Requirements: The project will be implemented in Python / C++ so good knowledge is required. Previous experience in one or several of the following topics would be a plus: image processing, computer vision or machine learning.

Assistant: Christophe Ecabert (christophe.ecabert@epfl.ch)

Supervisor: Prof. Jean-Philippe Thiran

11. Apizoom – deep learning to quantify the Verroa parasite in honey bee hive images (can be considered as a Master project in industry)

Varroa mites are recognized as the biggest pest to honey bees worldwide, and are believed to be the single largest contributing factor in the modern-day decline of honey bees due to their ability to transmit diseases, resulting in death or severe deformity of the pupae.

Verroa on honey bees.

Detecting and quantifying the presence of Verroa in a beehive is therefore crucial to treat the infection appropriately and as early as possible, and image analysis appears very useful in this problem.

In this project, we propose to develop an image analysis to detect and count Verroa cadavers who felt down on a plate below the beehive, as a non-intrusive way to quantify the presence of the Verroa. High definition images will be capture and Deep Learning techniques will be investigated here, to design and train a Convolutional Neural Network (CNN) to detect the Verroa and distinguish it from other wastes. See the following video for more details on the project (in French).

Depending on the evolution of the project, several steps will be investigated:

  • Development of the Deep Learning method
  • Training on a collection of annotated images
  • Test and improve

And possibly:

  • Study integration in mobile phones
  • Study the development of a web-based analysis system.

This project is jointly proposed with the company Apizoom (Fribourg, Switzerland).

Responsible: Prof. J.-Ph. Thiran

12. CleanCityIndex – A Deep Learning based system to localize and classify wastes on the streets

A review of major European cities places “urban cleanliness” as a top priority for the authorities, as it directly impacts the concern and satisfaction of their citizens and the attractiveness of their economy and tourism. Littering quantification is an important step toward improving urban cleanliness. When human interpretation is too cumbersome or in some cases impossible, an objective index of cleanliness could reduce the littering by awareness actions.

The goal of this project is to propose a fully automated computer vision application for littering quantification, based on images taken from the streets and sidewalks. We employ a deep learning based framework to localize and classify different types of wastes.

In this project, the student will be involved into study and develop new detection algorithms and investigates deep learning based object tracking techniques for this specific case. The student is expected to be familiar with Python and Tensorflow.

Assistant: Saeed Rad (saeed.rad@epfl.ch)
Supervisor: Prof. Jean-Philippe Thiran