Student Projects

LTS5 OPEN SEMESTER AND MASTER PROJECTS – Spring 2022

SEMESTER PROJECT PROPOSALS

1. Spherical Deconvolution Algorithms for Intra-Voxel Fiber Estimation and Brain Connectivity Mapping
The LTS5 Diffusion group focuses on brain tissue microstructure and structural connectivity –estimated by diffusion Magnetic Resonance Imaging (dMRI) data, with a particular focus on the reconstruction of the nerve fiber orientation distribution function (ODF) per voxel (see the figure below). This information is important for the reconstruction of the brain’s white matter by using fiber tracking algorithms (see ref [1]).

 

We have implemented various novel reconstruction algorithms (e.g., see refs. [2-5]) and we plan to develop a new generation of methods using Machine Learning techniques. The goals of this project are: (1) create a large database of fiber ODFs and corresponding dMRI signals, (2) Design, train, and optimize a neural network using this dataset, (3) predict the fiber ODFs from new dMRI data, and (4) compare the implemented algorithm with state-of-the-art techniques using both synthetic and real dMRI data acquired from human brains. The results will be published in international conferences and relevant journals.

Requirements: The project will be implemented in Python, so good knowledge is required. This project is ideal for a computer scientist, mathematician, physicist, or engineer interested in medical imaging, machine learning, signal processing, and optimization.

References:
[1] https://www.sciencedirect.com/science/article/abs/pii/S1053811914003541

[2] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4607500
[3] https://www.sciencedirect.com/science/article/abs/pii/S1053811918307699
[4] https://onlinelibrary.wiley.com/doi/10.1002/mrm.21917

Supervisors: Dr. Erick J. Canales-Rodríguez ([email protected]), Dr. Gabriel Girard ([email protected]), and Prof. Jean-Philippe Thiran


2. Myelin Water Imaging Using T2 Relaxometry
Myelin is a lipid-rich substance that surrounds the axons in the brain, which is essential for the proper functioning of the nervous system. Myelin water imaging is a magnetic resonance imaging (MRI) method that can be used to quantify and visualize myelination in the brain and spinal cord in vivo. The signal coming from the MRI machine (using a multi-echo T2 relaxometry sequence) can be decomposed into components, including that originated by water molecules trapped between the lipid bilayers of myelin. The correct estimation of this component provides a myelin-specific MRI biomarker to monitor brain changes in cerebral white matter. Myelin quantification has important implications for understanding various neurodegenerative diseases, including multiple sclerosis.
We are looking for a motivated student to (1) learn about the MRI and signal processing theory behind this modality, (2) improve the current estimation methods, and (3) test and compare the new results with the current methods and histological measurements.
The project builds on top of previous cutting-edge research carried out in our lab (for more details see our multi-component T2 reconstruction toolbox and related references: https://github.com/ejcanalesr/multicomponent-T2-toolbox). The results will be published in international conferences and relevant journals.

Requirements: The project will be implemented in Python, so good knowledge is required. This project is ideal for a computer scientist, mathematician, physicist, or engineer interested in medical imaging, optimization, and signal processing.

Supervisors: Dr. Erick J. Canales-Rodríguez ([email protected]) and Prof. Jean-Philippe Thiran


3. Debiased Scene Representation for Self-Supervised Learning

 Self-supervised learning, particularly contrastive learning, has demonstrated its strong capability to learn visual representations from unlabeled data. However, the learned representations are often biased to the spurious scene correlations caused by semantical different objects or backgrounds. This project aims to tackle this issue and formulate advanced self-supervised techniques via proper data augmentations considering the inferred object locations.

Reference:

[1] R. R. Selvaraju, K. Desai, J. Johnson, and N. Naik. Casting your model: Learning to localize improves self-supervised representations. In IEEE Conference on Computer Vision and Pattern Recognition, 2021.

[2] Mo, S., Kang, H., Sohn, K., Li, C.L. and Shin, J., 2021. Object-aware Contrastive Learning for Debiased Scene Representation. arXiv preprint arXiv:2108.00049.

Prerequisite: fluent in Pytorch and Python, familiar with deep learning fundamentalsAssistant: Dr Behzad Bozorgtabar ([email protected])

Supervisor: Prof. Jean-Philippe Thiran


4. Self-Supervised Learning-Based Domain Generalization and OoD Detection

Deep learning methods demonstrate their strong capability to handle independent and identically distributed (i.i.d) data in an experimental environment. Still, they often suffer from out-of-distribution (OoD) generalization, where the test data come from unseen visual domains with different distributions, w.r.t. the training one. Thus, deep neural networks are inherently restricted by the visual and semantic information available in their training data. Besides, collecting annotated data of all possible visual domains is often expensive or not feasible. This project aims to formulate new optimization and design deep architectures to learn a model that can generalize to previously unseen visual domains and recognize novel semantic concepts. More importantly, motivated by the success of self-supervised learning, particularly contrastive learning, the project will investigate a model to learn domain-invariant representation and new semantic augmentation techniques.

Reference:

 [1] Massimiliano Mancini, Zeynep Akata, Elisa Ricci, and Barbara Caputo. “Towards recognizing unseen categories in unseen domains”. In:proceedings of the European Conference on Computer Vision. Springer. 2020

[2] Kaiyang  Zhou, Yongxin  Yang, Yu  Qiao, and  Tao  Xiang.  “Domain  Generalization  with MixStyle”. In: International Conference on Learning Representations. 2021.

Prerequisite: fluent in Pytorch and Python, familiar with deep learning fundamentals

Assistant: Dr Behzad Bozorgtabar ([email protected])

Supervisor: Prof. Jean-Philippe Thiran


5. Dynamic image analysis for Cervical Cancer Detection – Analyse d’images dynamiques pour la détection du cancer du col de l’utérus

Cervical cancer is a major concern in public health around the world, both in high and low- and middle-income settings. In collaboration with the Geneva University Hospitals (HUG) and Dschang District Hospital in Cameroon, we aim at implementing a smartphone-based solution that automatically detects cervical cancer from videos of the cervix using deep neural networks.  

A first version of the algorithm has been implemented and requires additional features on two main aspects, each being subject to a student project. The opportunities are the following: 

  • Movement detection: The goal is to develop an algorithm detecting movement while the 2 minutes long video is taken by the doctors or midwives. This feature is essential to assess the images quality before performing their analysis. 

  • Region of interest detection: The goal is to automatically detect the location of the cervix in an image. This feature is essential for the algorithm to process only the region of interest and to avoid human intervention for identifying the location of the cervix. 

Other projects can possibly be conducted for identifying blood or mucus on the cervical images or the type of transformation zone (TZ) relying on deep learning. 

These projects will be realized in collaboration with the Department of Gynecology of the Geneva University Hospital.

Requirements: Python, deep learning, pytorch

Assistants: Roser Vinals Terres ([email protected]) et Magali Cattin ([email protected]

Supervisor: Prof. Jean-Philippe Thiran


6. Deep learning for automatic segmentation of tumour-affected kidney CT-scan

The nephroblastoma is one of the most frequently encountered tumours in children and accounts for between 5 to 12% of pediatric malignancies. A precise visualization of the tumour, the renal cortex, the excretory cavities and the vascular structures would help physicians when making decisions about which therapy to initiate and increase the safety of removal surgery.

The 3D representation of a tumour-affected kidney requires segmentation of the corresponding scanner, which is a time consuming and error prone task when performed manually. Developing an automatic segmentation method would make the approach more practicable.

Artificial intelligence, in particular Deep Learning (DL), is a branch of data analysis that is known to perform well at several computer vision tasks, including medical image segmentation. Unfortunately, the small amount of available data and the inter-patient variability render a naive DL method inadequate. This issue must be countered by combining DL with pre- and post-processing techniques relying on prior knowledge about the imaging method and the anatomy involved.

In this project, the student will use a state-of-the-art deep-learning model and improve its straightforward performance with hyper-parameter tuning and suitable image processing steps. The student is expected to be familiar with Python and Pytorch.

Assistant: Sandra Marcadent ([email protected])
Supervisor: Prof. Jean-Philippe Thiran


7. Computer vision for the EPFL Racing Team

EPFL Racing team is a student team that competes in the Formula Student competition, the largest student competition for engineers in the world. We design and build electric race cars and compete against hundreds of other universities during summer, all across Europe.

Teams are scored on different ‘events’ such as car acceleration, endurance, agility and are driven by pilots that are students, members of the team. Performance wise, to give an example the 0 – 100 km/h is realized in approximately 2.8 sec!

Starting in summer 2021 we plan on having an autonomous car, able to visualize the track and to find the perfect line to get the best lap time! To do so, we are looking to start projects on the topics of computer vision.

The aim of the projects would be to find and implement the most performant and robust detection algorithms, knowing that the track is bounded by yellow and blue cones (cf. picture of the car of ETH in 2017). The car must be as quick as possible around the track while hitting as few cones as possible, otherwise we get a time penalty. Also, it will be crucial to consider the hardware implementation on the 2021 car: what sensors, where to put them and how much power is required.

Supervisor: Prof. Jean-Philippe Thiran


8. 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.

Supervisor: Prof. Jean-Philippe Thiran



MASTER PROJECT PROPOSALS 

MEDICAL IMAGING PROJECTS

1. Spherical Deconvolution Algorithms for Intra-Voxel Fiber Estimation and Brain Connectivity Mapping
The LTS5 Diffusion group focuses on brain tissue microstructure and structural connectivity –estimated by diffusion Magnetic Resonance Imaging (dMRI) data, with a particular focus on the reconstruction of the nerve fiber orientation distribution function (ODF) per voxel (see the figure below). This information is important for the reconstruction of the brain’s white matter by using fiber tracking algorithms (see ref [1]).

 

We have implemented various novel reconstruction algorithms (e.g., see refs. [2-5]) and we plan to develop a new generation of methods using Machine Learning techniques. The goals of this project are: (1) create a large database of fiber ODFs and corresponding dMRI signals, (2) Design, train, and optimize a neural network using this dataset, (3) predict the fiber ODFs from new dMRI data, and (4) compare the implemented algorithm with state-of-the-art techniques using both synthetic and real dMRI data acquired from human brains. The results will be published in international conferences and relevant journals.

Requirements: The project will be implemented in Python, so good knowledge is required. This project is ideal for a computer scientist, mathematician, physicist, or engineer interested in medical imaging, machine learning, signal processing, and optimization.

References:
[1] https://www.sciencedirect.com/science/article/abs/pii/S1053811914003541

[2] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4607500
[3] https://www.sciencedirect.com/science/article/abs/pii/S1053811918307699
[4] https://onlinelibrary.wiley.com/doi/10.1002/mrm.21917

Supervisors: Dr. Erick J. Canales-Rodríguez ([email protected]), Dr. Gabriel Girard ([email protected]), and Prof. Jean-Philippe Thiran


2. Myelin Water Imaging Using T2 Relaxometry
Myelin is a lipid-rich substance that surrounds the axons in the brain, which is essential for the proper functioning of the nervous system. Myelin water imaging is a magnetic resonance imaging (MRI) method that can be used to quantify and visualize myelination in the brain and spinal cord in vivo. The signal coming from the MRI machine (using a multi-echo T2 relaxometry sequence) can be decomposed into components, including that originated by water molecules trapped between the lipid bilayers of myelin. The correct estimation of this component provides a myelin-specific MRI biomarker to monitor brain changes in cerebral white matter. Myelin quantification has important implications for understanding various neurodegenerative diseases, including multiple sclerosis.
We are looking for a motivated student to (1) learn about the MRI and signal processing theory behind this modality, (2) improve the current estimation methods, and (3) test and compare the new results with the current methods and histological measurements.
The project builds on top of previous cutting-edge research carried out in our lab (for more details see our multi-component T2 reconstruction toolbox and related references: https://github.com/ejcanalesr/multicomponent-T2-toolbox). The results will be published in international conferences and relevant journals.

Requirements: The project will be implemented in Python, so good knowledge is required. This project is ideal for a computer scientist, mathematician, physicist, or engineer interested in medical imaging, optimization, and signal processing.

Supervisors: Dr. Erick J. Canales-Rodríguez ([email protected]) and Prof. Jean-Philippe Thiran


3. Dynamic image analysis for Cervical Cancer Detection – Analyse d’images dynamiques pour la détection du cancer du col de l’utérus

Cervical cancer is a major concern in public health around the world, both in high and low- and middle-income settings. In collaboration with the Geneva University Hospitals (HUG) and Dschang District Hospital in Cameroon, we aim at implementing a smartphone-based solution that automatically detects cervical cancer from videos of the cervix using deep neural networks.  

A first version of the algorithm has been implemented and requires additional features on two main aspects, each being subject to a student project. The opportunities are the following: 

  • Movement detection: The goal is to develop an algorithm detecting movement while the 2 minutes long video is taken by the doctors or midwives. This feature is essential to assess the images quality before performing their analysis. 

  • Region of interest detection: The goal is to automatically detect the location of the cervix in an image. This feature is essential for the algorithm to process only the region of interest and to avoid human intervention for identifying the location of the cervix. 

Other projects can possibly be conducted for identifying blood or mucus on the cervical images or the type of transformation zone (TZ) relying on deep learning. 

These projects will be realized in collaboration with the Department of Gynecology of the Geneva University Hospital.

Requirements: Python, deep learning, pytorch

Assistants: Roser Vinals Terres ([email protected]) et Magali Cattin ([email protected]

Supervisor: Prof. Jean-Philippe Thiran


4. Self-supervised learning for computational pathology and survival prediction.

Computational pathology is a state-of-the-art technology that relies on histopathological images to predict a large variety of components such as tumor regions, clinical scores, or patient overall survival. Standard supervised machine learning approaches for tissue classifications and representations are limited by the amount of labeled data which are often rare in the medical field. Moreover, they often struggle to handle, in an efficient way, the large amount of data generated by the scanners.

Self-supervised learning is a branch of unsupervised learning that takes advantage of the structure of the data itself to learn meaningful features representations and, therefore, does not rely on labeled data. In addition, as the supervised aspect of the learning process directly comes from the data themselves, the models can benefit from the large quantity of generated data.

In this project, the student will use state-of-the-art self-supervised models to learn discriminative features for several medical tasks such as cancer region detection and patient survival prediction. The student is expected to be familiar with Python and Pytorch.

Assistant: Christian Abbet ([email protected])
Supervisor: Prof. Jean-Philippe Thiran


5. Digital histopathology slide processing using deep learning methods and its applications
Recently, the rise of many novel digital histopathology imaging systems promises to revolutionize pathologist’s workflow and create opportunities for the application of image processing & machine learning methods in both translational research and diagnostic pathology. These methods aim to improve the accuracy and reproducibility of pathologists’ interpretations by limiting inter and intra-observer variability and ideally providing more precise information for outcome prediction. In practice, automatic and effective tissue segmentation and classification are often crucial steps for the success of computational pathology pipelines. Benefiting the existing resource of big cohorts and different image processing methods developed in the Institute of Pathology (UniBe), this project targets to improve the performance of these methods (i.e: tissue segmentation & classification). Then, the student will build the application to make these methods more accessible for the pathologists and researchers in daily tasks. The student is expected to be familiar with Python and Pytorch.

Assistant: Christian Abbet ([email protected]), Dr. Giao Nguyen ([email protected]),
Supervisor: Prof. Jean-Philippe Thiran, Prof. Inti Zlobec (Uni Bern).


6. Deep learning for automatic segmentation of tumour-affected kidney CT-scan

The nephroblastoma is one of the most frequently encountered tumours in children and accounts for between 5 to 12% of pediatric malignancies. A precise visualization of the tumour, the renal cortex, the excretory cavities and the vascular structures would help physicians when making decisions about which therapy to initiate and increase the safety of removal surgery.

The 3D representation of a tumour-affected kidney requires segmentation of the corresponding scanner, which is a time consuming and error prone task when performed manually. Developing an automatic segmentation method would make the approach more practicable.

Artificial intelligence, in particular Deep Learning (DL), is a branch of data analysis that is known to perform well at several computer vision tasks, including medical image segmentation. Unfortunately, the small amount of available data and the inter-patient variability render a naive DL method inadequate. This issue must be countered by combining DL with pre- and post-processing techniques relying on prior knowledge about the imaging method and the anatomy involved.

In this project, the student will use a state-of-the-art deep-learning model and improve its straightforward performance with hyper-parameter tuning and suitable image processing steps. The student is expected to be familiar with Python and Pytorch.

Assistant: Sandra Marcadent (sandr[email protected])
Supervisor: Prof. Jean-Philippe Thiran


7. Deep learning for privacy-preserving medical imaging of the head: anonymous face generation and adversarial attacks

Description and Objective
Magnetic Resonance Imaging (MRI) is a well‐established medical imaging modality exhibiting excellent soft tissue contrast which makes it the method of choice for brain structure assessment. Routine structural brain MRI scans allow 3D face-reconstruction of individuals and can thus enable subject identification [Schwarz et al., N Engl J Med, 2019]. This poses a problem for privacy, particularly when such data is being shared or made publicly available which has led to a demand to de-identify the actual MR image data e.g. by defacing or blurring. Such image de-identification can, however, pose a challenge for commonly used automated post-processing techniques like brain volume assessment or lesion detection that are often designed to have normal whole-head MR datasets as input. Using such algorithms with de-identified data can affect results or even cause complete failure [de Sitter et al., European Radiology, 2020]. To address this challenge, different methods have recently been proposed to, instead of removing facial structures, replace them with generic or artificially generated ones that are not related to the original facial features and therefore do not allow subject identification. Despite these efforts, many problems are still unsolved such as limitation to T1-weighted MRI data or consistency for follow-up exams. Finding solutions for these challenges is highly relevant for future medical image processing applications and related research.
The goal of this master thesis project is to develop a robust method for de-identifying head MRI scans while keeping the impact of this process on existing post-processing solutions as small as possible. The method should work irrespective of the used MR image contrast and provide consistent results across different image contrast as well as for acquisitions from different systems or at different time points. To achieve this, novel approaches such as combination of state-of-the-art machine learning techniques with parametric face shape models will be explored. Further to this a built-in quality control mechanism needs to be developed to ensure correct de-identification as well as maintained post-processability. A thorough evaluation that these requirements are met is an essential part of the project and will involve large real-world MRI datasets together with current post-processing applications. In addition, we will develop adversarial attacks that attempt to re-identify specific subjects from anonymized images, and test conditions and necessary data for the attacks to succeed.


Required skills
Background in computer science, electrical engineering, biomedical engineering, or similar. Previous programming experience in Python.

Supervisor
Prof. Jean-Philippe Thiran (EPFL-LTS5)

Co-supervisors
Dr Till Hulnhagen (SIEMENS HEALTHINEERS)
Dr Jonas Richiardi (CHUV)


8. AI-based Predictive Model of minimally invasive intervention of liver tumors: First Large-scale Clinical Evaluation

As of 2017, cancer is the second deadliest non-communicable disease (NCD) around the world after cardio-vascular disease and before chronic respiratory disease and diabetes. Among the different types of cancer, liver cancer is the fourth deadliest cancer in the world and its overall death rate has more than doubled from 1980 to 2016 in the United States [Naghavi et al. The Lancet, 2017].

Thermal ablation (TA) is a minimally invasive alternative to treat liver cancer. It is far less invasive, has lower complication rates, a superior cost-effectiveness, and an extremely lower treatment-associated mortality than surgery [Heimbach et al. Hepatology 2018]. Despite a very favorable safety profile and a curative potential, TA still suffers from a relatively high local tumor progression (LTP) rates compared to surgery. Treatment failure and LTP can be attributed to an insufficient coverage of the tumor by the ablation zone: tumor and minimal ablative margin (MAM).

Therefore, an accurate and reliable prediction of the ablation in real time that could estimate the a-priori MAM remains one of the missing pieces to reduce the number of unexpected ablation results. However, it is still challenging to accurately predict the overall ablation volume and shape, mostly due to the heat sink effect of the large hepatic blood vessels.

Computational models of thermal ablation have been proposed to improve the accuracy of the ablation zone prediction. RF manufacturers usually model the ablation as an ellipsoid or a sphere of radius ranging from 1 to 4 cm based on the type of tissue being ablated as well as the ablation power and duration. More advanced computational models simulate the heat diffusion in the liver, predict the temperature distribution during the procedure and finally evaluate the ablation zone. Several patient-specific models of heat transfer in biological tissue have been proposed to include patient-specific parameters like vascular anatomy and hepatic perfusion in order to improve the accuracy of the ablation zone prediction.

The goal of this master thesis project is to develop a robust and fast framework for the clinical validation of a predictive computational model of thermal ablation. A thorough evaluation of the method is an essential part of the project and will involve large real-world datasets. This project will be performed in close collaboration with the interventional radiology department of the CHUV to always ensure a relevant clinical application.

Required skills
Background in computer science, electrical engineering, biomedical engineering, or similar.
Programming experience in Python and/or C++.

Supervisor
Prof Jean-Philippe Thiran (LTS5)

Co-supervisors
Dr Chloé Audigier (SIEMENS HEALTHINEERS)


COMPUTER VISION PROJECTS ​​

9. Debiased Scene Representation for Self-Supervised Learning

 Self-supervised learning, particularly contrastive learning, has demonstrated its strong capability to learn visual representations from unlabeled data. However, the learned representations are often biased to the spurious scene correlations caused by semantical different objects or backgrounds. This project aims to tackle this issue and formulate advanced self-supervised techniques via proper data augmentations considering the inferred object locations.

Reference:

[1] R. R. Selvaraju, K. Desai, J. Johnson, and N. Naik. Casting your model: Learning to localize improves self-supervised representations. In IEEE Conference on Computer Vision and Pattern Recognition, 2021.

[2] Mo, S., Kang, H., Sohn, K., Li, C.L. and Shin, J., 2021. Object-aware Contrastive Learning for Debiased Scene Representation. arXiv preprint arXiv:2108.00049.

Prerequisite: fluent in Pytorch and Python, familiar with deep learning fundamentalsAssistant: Dr Behzad Bozorgtabar ([email protected])

Supervisor: Prof. Jean-Philippe Thiran


10. Self-Supervised Learning-Based Domain Generalization and OoD Detection

Deep learning methods demonstrate their strong capability to handle independent and identically distributed (i.i.d) data in an experimental environment. Still, they often suffer from out-of-distribution (OoD) generalization, where the test data come from unseen visual domains with different distributions, w.r.t. the training one. Thus, deep neural networks are inherently restricted by the visual and semantic information available in their training data. Besides, collecting annotated data of all possible visual domains is often expensive or not feasible. This project aims to formulate new optimization and design deep architectures to learn a model that can generalize to previously unseen visual domains and recognize novel semantic concepts. More importantly, motivated by the success of self-supervised learning, particularly contrastive learning, the project will investigate a model to learn domain-invariant representation and new semantic augmentation techniques.

Reference:

 [1] Massimiliano Mancini, Zeynep Akata, Elisa Ricci, and Barbara Caputo. “Towards recognizing unseen categories in unseen domains”. In:proceedings of the European Conference on Computer Vision. Springer. 2020

[2] Kaiyang  Zhou, Yongxin  Yang, Yu  Qiao, and  Tao  Xiang.  “Domain  Generalization  with MixStyle”. In: International Conference on Learning Representations. 2021.

Prerequisite: fluent in Pytorch and Python, familiar with deep learning fundamentals

Assistant: Dr Behzad Bozorgtabar ([email protected])

Supervisor: Prof. Jean-Philippe Thiran


11. 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.

Supervisor: Prof. Jean-Philippe Thiran


12. Computer vision for the EPFL Racing Team

EPFL Racing team is a student team that competes in the Formula Student competition, the largest student competition for engineers in the world. We design and build electric race cars and compete against hundreds of other universities during summer, all across Europe.

Teams are scored on different ‘events’ such as car acceleration, endurance, agility and are driven by pilots that are students, members of the team. Performance wise, to give an example the 0 – 100 km/h is realized in approximately 2.8 sec!

Starting in summer 2021 we plan on having an autonomous car, able to visualize the track and to find the perfect line to get the best lap time! To do so, we are looking to start projects on the topics of computer vision.

The aim of the projects would be to find and implement the most performant and robust detection algorithms, knowing that the track is bounded by yellow and blue cones (cf. picture of the car of ETH in 2017). The car must be as quick as possible around the track while hitting as few cones as possible, otherwise we get a time penalty. Also, it will be crucial to consider the hardware implementation on the 2021 car: what sensors, where to put them and how much power is required.

Supervisor: Prof. Jean-Philippe Thiran


13. Master thesis/intership with Nestlé Research, Nestlé Institute of Health Science

See attached file NestléResearchLausanne MSc Placement 2022.