Student Projects

LTS5 OPEN SEMESTER AND MASTER PROJECTS – Spring 2021

SEMESTER PROJECT PROPOSALS

1. Multiview 3D Face Reconstruction

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.

Conventional 3D reconstruction technics work on single image with the help of explicit models of the object (i.e. Face) due to the lack of depth information. Working with only one image brings challenges dues to external factors such as occlusions, lighting and large head pose. Therefore having multiple view angles of the same object can simplify and increase the robustness of the system.

The goals of this project are in two folds. Develop a solution to extend current 3D reconstruction framework to work with multiple source images. Finally benchmark both approaches and compare their performance in a quantitative study.

Requirements: The project will be implemented in Python 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 ([email protected])

Supervisor: Prof. Jean-Philippe Thiran

2. 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).

Assistant: Roser Vidals Terres ([email protected])

Responsible: Prof. J.-Ph. Thiran

3. 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

4. Spherical Deconvolution Algorithms for Brain Connectivity Mapping

The LTS5 Diffusion group focuses on structural brain connectivity – estimated by diffusion Magnetic Resonance Imaging, with a particular focus on the local reconstruction of nerve fibers orientations and tractography algorithms (see the figure below).

 

Tractography reconstruction of the brain’s white matter. The figure shows millions of 3D lines approximating the trajectories of bundles of axons interconnecting the brain.

We have implemented various novel reconstruction algorithms and we plan to include some of these tools into the open-source Diffusion Imaging in Python Library (DIPY: https://dipy.org/). The goals of this project are: (1) translate our current Matlab implementations to Python3, (2) accelerate and parallelize the estimation process (CPU/GPU), (3) study the accuracy and speed of the implemented methods, and (4) publish the final version on DIPY.

Requirements: The project will be implemented in Python/Matlab so good knowledge is required (experience in C++ is a plus). This project is ideal for a computer scientist or engineer interested in signal processing, code optimization, and medical imaging.

References: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4607500/

https://www.sciencedirect.com/science/article/abs/pii/S1053811918307699

https://www.sciencedirect.com/science/article/abs/pii/S1053811914003541

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

5. Abnormality Detection for X-Ray Radiographic Images

The X-ray radiograph (XR) such as chest X-ray is frequently acquired for detecting and diagnosing pathologies and it is widely used in cancer prevention and treatment planning. However, the radiologist reporting of every image is a tedious and time-consuming task. Having said that, abnormal cases are usually much rarer than normal cases . In addition, it is very time-consuming to collect a reasonable amount for abnormal cases, while on the other hand, normal XRs are much easier to obtain. Therefore, fully supervised deep neural networks’ (DNNs) effectiveness is reduced. Consequently, an automatic system of XR abnormality detection would be advantageous, allowing reducing time for annotation and paying more attention on pathology analysis of abnormal XRs. The goal of this project is to develop a computer-aided diagnosis (CAD) system that can automatically identify abnormal XRs by learning only from normal ones during training phase.

In this project, the student will be involved in study and develop DNN-based one-class classifier for abnormal X-ray detection. The student is expected to be familiar with Python and Pytorch.

Assistant: Dr Behzad Bozorgtabar ([email protected])

Supervisor: Prof. Jean-Philippe Thiran

6. Image Modality Conversion

 Recently, interest in using image modality conversion for radiotherapy applications have grown rapidly. For example, for replacing Computed tomography (CT) with magnetic resonance imaging (MRI) for diagnostic and therapeutic purposes. This is mainly because MRI delivers superior contrast of soft tissue compared with the CT scans and is free of ionizing radiation. Therefore, modality conversion is desirable for many applications such as CT-based radiotherapy. Recent studies have shown remarkable success in image-to-image translation for modality conversion. This task has experienced significant improve- ments following the introduction of generative adversarial networks (GANs).

In this project, the student will be involved in study and develop a DNN-based image synthesis approach for wide range of modality conversion with the focus on radiotherapy applications. The student is expected to be familiar with Python and Pytorch.

Assistant: Dr Behzad Bozorgtabar ([email protected])

Supervisor: Prof. Jean-Philippe Thiran

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: Python, image processing, skills in development for Android would be ideal.

Assistant: Roser Vidals Terres ([email protected])

Supervisor: Prof. Jean-Philippe Thiran

8. Automated splitting of multiple sclerosis lesions in magnetic resonance imaging

Total lesion burden (or total lesion load) based on magnetic resonance imaging (MRI) analysis is a widely recognized biomarker in multiple sclerosis (MS), however, its associations with the clinical indicators of disease severity have always shown to be limited. On the other hand, recent studies have shown that the distinct number of lesions in a patient’s brain might correlate with the Expanded Disability Status Scale (EDSS). Furthermore, an accurate localization of the distinct lesions would allow to apply novel machine learning techniques which require separate connected components per each lesion. However, due to the numerous confluent lesions appearing during the pathological progression of the disease, estimating the exact number of lesions present is challenging and time-consuming even for experts. Moreover, today there is not a reliable automatic method to speed-up the process.

This project aims at developing an automated method for the assessment of MS confluent lesions from MRI. Based on recent research works, this project will further investigate the combination of morphometrical, lesion probability, and texture features observed in different MRI contrasts with the aim of better analyzing and splitting confluent lesions.

Image from Dworkin, Jordan D., et al. “An automated statistical technique for counting distinct multiple sclerosis lesions.” AJN 2018.

Required skills: Background in physics, computer science, electrical engineering, biomedical engineering, or similar. Basic knowledge of image processing methods. Good knowledge of Python. Experience with MRI as well as medical imaging processing pipelines (ANTs, FreeSurfer,..) is a plus.

Assistant: Francesco La Rosa ([email protected])

Supervisors: Merixtell Bach Cuadra, Jean-Philippe Thiran

9. 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

 

MASTER PROJECT PROPOSALS 

MEDICAL IMAGING PROJECTS

1. Abnormality Detection for X-Ray Radiographic Images

The X-ray radiograph (XR) such as chest X-ray is frequently acquired for detecting and diagnosing pathologies and it is widely used in cancer prevention and treatment planning. However, the radiologist reporting of every image is a tedious and time-consuming task. Having said that, abnormal cases are usually much rarer than normal cases . In addition, it is very time-consuming to collect a reasonable amount for abnormal cases, while on the other hand, normal XRs are much easier to obtain. Therefore, fully supervised deep neural networks’ (DNNs) effectiveness is reduced. Consequently, an automatic system of XR abnormality detection would be advantageous, allowing reducing time for annotation and paying more attention on pathology analysis of abnormal XRs. The goal of this project is to develop a computer-aided diagnosis (CAD) system that can automatically identify abnormal XRs by learning only from normal ones during training phase.

In this project, the student will be involved in study and develop DNN-based one-class classifier for abnormal X-ray detection. The student is expected to be familiar with Python and Pytorch.

Assistant: Dr Behzad Bozorgtabar ([email protected])

Supervisor: Prof. Jean-Philippe Thiran

2. Image Modality Conversion

 Recently, interest in using image modality conversion for radiotherapy applications have grown rapidly. For example, for replacing Computed tomography (CT) with magnetic resonance imaging (MRI) for diagnostic and therapeutic purposes. This is mainly because MRI delivers superior contrast of soft tissue compared with the CT scans and is free of ionizing radiation. Therefore, modality conversion is desirable for many applications such as CT-based radiotherapy. Recent studies have shown remarkable success in image-to-image translation for modality conversion. This task has experienced significant improve- ments following the introduction of generative adversarial networks (GANs).

In this project, the student will be involved in study and develop a DNN-based image synthesis approach for wide range of modality conversion with the focus on radiotherapy applications. The student is expected to be familiar with Python and Pytorch.

Assistant: Dr Behzad Bozorgtabar ([email protected])

Supervisor: Prof. Jean-Philippe Thiran

3. 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: Python, image processing, skills in development for Android would be ideal.

Assistant: Roser Vidals Terres ([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. Automated splitting of multiple sclerosis lesions in magnetic resonance imaging

Total lesion burden (or total lesion load) based on magnetic resonance imaging (MRI) analysis is a widely recognized biomarker in multiple sclerosis (MS), however, its associations with the clinical indicators of disease severity have always shown to be limited. On the other hand, recent studies have shown that the distinct number of lesions in a patient’s brain might correlate with the Expanded Disability Status Scale (EDSS). Furthermore, an accurate localization of the distinct lesions would allow to apply novel machine learning techniques which require separate connected components per each lesion. However, due to the numerous confluent lesions appearing during the pathological progression of the disease, estimating the exact number of lesions present is challenging and time-consuming even for experts. Moreover, today there is not a reliable automatic method to speed-up the process.

This project aims at developing an automated method for the assessment of MS confluent lesions from MRI. Based on recent research works, this project will further investigate the combination of morphometrical, lesion probability, and texture features observed in different MRI contrasts with the aim of better analyzing and splitting confluent lesions.

Image from Dworkin, Jordan D., et al. “An automated statistical technique for counting distinct multiple sclerosis lesions.” AJN 2018.

Required skills: Background in physics, computer science, electrical engineering, biomedical engineering, or similar. Basic knowledge of image processing methods. Good knowledge of Python. Experience with MRI as well as medical imaging processing pipelines (ANTs, FreeSurfer,..) is a plus.

Assistant: Francesco La Rosa ([email protected])

Supervisors: Merixtell Bach Cuadra, Jean-Philippe Thiran

6. Deep learning for volumetric time series segmentation – towards fast and accurate liver cancer screening from clinical Magnetic Resonance Images

Supervisors: Jonas Richiardi, Naïk Vietti Violi (CHUV), Jean-Philippe Thiran (EPFL)

INTRODUCTION

Hepatocellular carcinoma (HCC) is the second leading cause of cancer-related death worldwide [Torre2012]. The stage of disease at time of diagnosis determines prognosis and treatment effectiveness. In at-risk populations, clinical guidelines recommend semi-annual screening with ultrasound (US). Unfortunately, US sensitivity is low, around 45% for early-stage HCC [Tzartzeva2018]. Magnetic Resonance Imaging (MRI) has higher sensitivity (80% or more [Colli2006]), but at ~45 minutes, full examinations are too long for clinical routine. Abbreviated MRI (AMRI) is a novel concept that consists in using a limited number of MRI sequences, thus reducing acquisition time (<10mins) and costs compared to MRI while keeping high diagnostic performance for HCC detection [Vietti Violi2020].

Here, our goal is to develop deep learning methods for HCC screening, using AMRI protocols, in order to improve lesion detection compared to human eye evaluation. One challenge is that the input MRI data is not only 3D (volumetric) but 4D (volumetric time series), and thus will require custom deep network design to obtain good performance for lesion detection.

This is a challenging translational machine-learning project with potentially large and rapid clinical impact, and requires prior experience with Python. Experience in deep learning and image processing is not required but would be an advantage.

APPROACH

We will approach the problem as a dense (semantic) segmentation task, where the goal is to segment liver lesions (3D volume) using T1-weighted Dynamic Contrast Enhanced MRI (4D volumetric time series). As a baseline architecture, we will use a U-net [Ronneberger2015], a well-performing and robust convolutional network, together with some common enhancements.

Our research work will then focus on finding the optimal representation for the 4D input data, and modifying the architecture to enable modeling of the temporal dependencies between the different time points (phases), and parameter sharing and regularization between the input kernels for the various phases and early feature maps.

Subsequently, we will explore the role of data heuristics developed on other segmentation tasks and open challenges that led to state-of-the-art performance [Isensee2020], as well as limited-budget neural architecture search (gradient-free optimization), and approach that we have previously seen to yield novel and high-accuracy network architectures.

DATA AND EVALUATION

We will use in-house data for 120 patients, each with 1-3 lesions, with dense annotations for both the liver and the lesions. In addition, we will use data from other clinical partners to assess generalization ability of the model. We will evaluate performance using lesion-level metrics (segmentation: Jaccard coefficient, absolute volume difference, detection: sensitivity, specificity, AUROC, AUPR) and patient-level metrics (sensitivity, specificity, PPV, NPV) for all-stage and early-stage HCC detection.

REFERENCES

Colli et al (2006) Accuracy of ultrasonography, spiral CT, magnetic resonance, and alpha-fetoprotein in diagnosing hepatocellular carcinoma: a systematic review, Am J Gastroenterol 101(3) (2006) 513-23.

Isensee et al (2020) Automated Design of Deep Learning Methods for Biomedical Image Segmentation, arXiv:1904.08128v2

Ronneberger et al. (2015) U-net: Convolutional networks for biomedical image segmentation, Proc. MICCAI 234-241

Vietti Violi et al (2020) Gadoxetate-enhanced abbreviated MRI is highly accurate for hepatocellular carcinoma screening, European Radiology

Torre et al (2015) Global cancer statistics, 2012. CA Cancer J Clin 65:87-108

Tzartzeva et al (2018) Surveillance Imaging and Alpha Fetoprotein for Early Detection of Hepatocellular Carcinoma in Patients With Cirrhosis: A Meta-analysis. Gastroenterology 154:1706-1718 e1701

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. Deep learning for 3D abdominal image registration

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]. Medical imaging can be used for liver cancer diagnostic, liver tumor detection as well as for treatment planning such as resection or thermal ablation. To investigate disease progression, or for an accurate treatment planning, there is a widespread interest in accurately relating the information in the different images of the same patient and this can be achieved with image registration. Image registration is the process of transforming different images into one common coordinate system. In the case of abdominal images, registration algorithms have to deal with large deformation that can be due to the breathing motion, or the interaction with a surgical device in the case of resection for example. Moreover, the liver can exhibit a highly different aspect if the two images are acquired before and after a surgery, or a thermal ablation.
The goal of this master thesis project is to develop a robust and fast registration method for abdomen images. To achieve this, novel approaches such as combination of state-of-the-art machine learning techniques will be explored. A thorough evaluation of the method is an essential part of the project and will involve large real-world datasets.

Supervisor : Prof. Jean-Philippe THIRAN

Co-supervisors : Dr Chloé Audigier (SIEMENS HEALTHINEERS)

This is a role well suited to an ambitious professional, looking for the next step in their career. As a M.Sc. student working on deep learning for 3D abdominal image registration, you will be responsible for:

  • Research, design, and implement machine learning algorithms for 3D image registration
  • Advance the state-of-the-art, including generating patents and top publications
  • Work on large-scale, real-world problems, be entrepreneurial in an environment that’s more startup than big company

This position may suit you best if you are familiar with what is below:

  • Current graduate students in Computer Science, Electrical Engineering, Applied Mathematics, Engineering or related disciplines for research interns
  • Experience with machine learning libraries such as PyTorch, TensorFlow, Keras
  • Familiarity with image processing in Python and C++
  • Hands-on coding skills and ability to quickly prototype in Python.
  • Knowledge of systems for software version control (SVN, Git) and software build (CMake) is a plus
  • Strong collaboration skills and ability to thrive in a fast-paced environment

 Required skills to have for the success of this role

  • Strong theoretical and practical background in machine learning and in particular deep learning
  • Strong written and verbal communication skills in English is required
  • Excellent interpersonal skills and a can-do attitude
  • Flexibility and adaptability to work in a growing, dynamic team

 

COMPUTER VISION PROJECTS

​​9. Multiview 3D Face Reconstruction

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.

Conventional 3D reconstruction technics work on single image with the help of explicit models of the object (i.e. Face) due to the lack of depth information. Working with only one image brings challenges dues to external factors such as occlusions, lighting and large head pose. Therefore having multiple view angles of the same object can simplify and increase the robustness of the system.

The goals of this project are in two folds. Develop a solution to extend current 3D reconstruction framework to work with multiple source images. Finally benchmark both approaches and compare their performance in a quantitative study.

Requirements: The project will be implemented in Python 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 ([email protected])

Supervisor: Prof. Jean-Philippe Thiran

10. 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).

Assistant: Roser Vidals Terres ([email protected])

Responsible: Prof. J.-Ph. 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