Theses

The primary goal of this thesis is to design and implement a user-friendly interface for video inpainting, allowing users to upload videos and remove unwanted objects or restore corrupted regions fast and easily. The user interface (UI) integrates two state-of-the-art models: a video object segmentation method used to mark the area for removal or restoration, and a robust inpainting technique that fills in that marked region and ensures fast and visually consistent results. Inpainting techniques have long been available for images, the next step in this field is to improve these methods for videos. While image inpainting can focus solely on achieving spatial consistency within a single image, video inpainting techniques have to also maintain temporal consistency across multiple frames. This significantly increases the computational complexity and processing time. A key challenge in this thesis is to find a compromise between reducing the runtime of these methods and keeping high-quality results. To overcome this, several experiments were conducted to optimize both video object segmentation and inpainting techniques.

This thesis presents Quasi-OmniFastTrack, an improved version of the OmniMotion algorithm for long-term pixel tracking in videos. The key contribution is reducing the computational expense and training time of OmniMotion while maintaining comparable tracking performance. The main bottleneck in OmniMotion was identified to be the NeRF network used for 3D scene representation. Quasi-OmniFastTrack replaces this with a pre-trained depth estimation model, significantly reducing training time, based on the work introduced in OmniFastTrack, hence the name. The invertible neural network for mapping between local and canonical coordinates is retained, but optimized depths are used to lift 2D pixels to 3D. Experiments show that Quasi-OmniFastTrack reduces training time by over 50% compared to OmniMotion while achieving similar qualitative tracking results on sequences with occlusions. Performance degrades somewhat on fast-moving scenes. The ablation studies demonstrate the importance of optimizing the initial depth estimates during training. While not matching OmniMotion's robustness in all scenarios, Quasi-OmniFastTrack offers a compelling speed-accuracy tradeoff, enabling long-term tracking on more videos in practical timeframes. Future work on incorporating other modifications introduced in OmniFastTrack, like long-term semantic features, could further improve tracking consistency.

This paper deals with exploring the importance of positional information learned in pre-training of a large language model like BERT. We introduce a simple pre-training task, predicting the absolute position of a given bag of words, and evaluate the impact on the performance of two downstream tasks: Named Entitiy Recognition and Sentiment Analysis. The challenge to the pre-training task is that all positional encoding has been removed and the model is required to learn the relationship between position of a word and its syntactic and semantic meaning in a sentence. In order to achive robuster and more natural results a simpler version of MLM as well as the Sinkhorn-Knopp Algorithm are explored. The positional information learned durning pre-training leads to a competitive understanding of language, which then leads to good results when transferred to downstream tasks. We further highlights some observations in relation with the setting of the pre-training task and evaluates the changes brought with the Sinkhorn-Knopp Algorithm and simple Masked Language Modeling.

The calculation of variables of brain morphology is computationally very expensive and time-consuming. A previous work showed the feasibility of ex- tracting the variables directly from T1-weighted brain MRI images using a con- volutional neural network. We used significantly more data and extended their model to a new set of neuromorphological variables, which could become inter- esting biomarkers in the future for the diagnosis of brain diseases. The model shows for nearly all subjects a less than 5% mean relative absolute error. This high relative accuracy can be attributed to the low morphological variance be- tween subjects and the ability of the model to predict the cortical atrophy age trend. The model however fails to capture all the variance in the data and shows large regional differences. We attribute these limitations in part to the moderate to poor reliability of the ground truth generated by FreeSurfer. We further investigated the effects of training data size and model complexity on this regression task and found that the size of the dataset had a significant impact on performance, while deeper models did not perform better. Lack of interpretability and dependence on a silver ground truth are the main drawbacks of this direct regression approach.

Detection and tracking of humans via UWB radars is a promising and continuously evolving field with great potential for medical technology. This contactless method of acquiring data of a patients movement patterns is ideal for in home application. As irregularities in a patients movement patterns are an indicator for various health problems including neurodegenerative diseases, the insight this data could provide may enable earlier detection of such problems. In this thesis a signal processing pipeline is presented with which a persons movement is modeled. During an experiment 142 measurements were recorded by two separate radar systems and one lidar system which each consisted of multiple sensors. The models that were calculated on these measurements by the signal processing pipeline were used to predict the times when a person stood up or sat down. The predictions showed an accuracy of 72.2%.

Arthroscopy consists of challenging tasks and requires skills that even today, young surgeons still train directly throughout the surgery. Existing simulators are expensive and rarely available. Through the growing potential of virtual reality(VR) (head-mounted) devices for simulation and their applicability in the medical context, these devices have become a promising alternative that would be orders of magnitude cheaper and could be made widely available. To build a VR-based training device for arthroscopy is the overall aim of our project, as this would be of great benefit and might even be applicable in other minimally invasive surgery (MIS). This thesis marks a first step of the project with its focus to explore and compare well-known algorithms in a multi-view stereo (MVS) based 3D reconstruction with respect to imagery acquired by an arthroscopic camera. Simultaneously with this reconstruction, we aim to gain essential measures to compare the VR environment to the real world, as validation of the realism of future VR tasks. We evaluate 3 different feature extraction algorithms with 3 different matching techniques and 2 different algorithms for the estimation of the fundamental (F) matrix. The evaluation of these 18 different setups is made with a reconstruction pipeline embedded in a jupyter notebook implemented in python based on common computer vision libraries and compared with imagery generated with a mobile phone as well as with the reconstruction results of state-of-the-art (SOTA) structure-from-motion (SfM) software COLMAP and Multi-View Environment (MVE). Our comparative analysis manifests the challenges of heavy distortion, the fish-eye shape and weak image quality of arthroscopic imagery, as all results are substantially worse using this data. However, there are huge differences regarding the different setups. Scale Invariant Feature Transform (SIFT) and Oriented FAST Rotated BRIEF (ORB) in combination with k-Nearest Neighbour (kNN) matching and Least Median of Squares (LMedS) present the most promising results. Overall, the 3D reconstruction pipeline is a useful tool to foster the process of gaining measurements from the arthroscopic exploration device and to complement the comparative research in this context.

In recent years deep convolutional neural networks achieved a lot of progress. To train such a network a lot of data is required and in supervised learning algorithms it is necessary that the data is labeled. To label data there is a lot of human work needed and this takes a lot of time and money to be done. To avoid the inconveniences that come with this we would like to find systems that don’t need labeled data and therefore are unsupervised learning algorithms. This is the importance of unsupervised algorithms, even though their outcome is not yet on the same qualitative level as supervised algorithms. In this thesis we will discuss an approach of such a system and compare the results to other papers. A deep convolutional neural network is trained to learn the rotations that have been applied to a picture. So we take a large amount of images and apply some simple rotations and the task of the network is to discover in which direction the image has been rotated. The data doesn’t need to be labeled to any category or anything else. As long as all the pictures are upside down we hope to find some high dimensional patterns for the network to learn.

This thesis explores the prospect of artificial neural networks for image processing tasks. More specifically, it aims to achieve the goal of stitching multiple overlapping images to form a bigger, panoramic picture. Until now, this task is solely approached with ”classical”, hardcoded algorithms while deep learning is at most used for specific subtasks. This thesis introduces a novel end-to-end neural network approach to image stitching called StitchNet, which uses a pre-trained autoencoder and deep convolutional networks. Additionally to presenting several new datasets for the task of supervised image stitching with each 120’000 training and 5’000 validation samples, this thesis also conducts various experiments with different kinds of existing networks designed for image superresolution and image segmentation adapted to the task of image stitching. StitchNet outperforms most of the adapted networks in both quantitative as well as qualitative results.

The idea of inferring the emotional state of a subject by looking at their face is nothing new. Neither is the idea of automating this process using computers. Researchers used to computationally extract handcrafted features from face images that had proven themselves to be effective and then used machine learning techniques to classify the facial expressions using these features. Recently, there has been a trend towards using deeplearning and especially Convolutional Neural Networks (CNNs) for the classification of these facial expressions. Researchers were able to achieve good results on images that were taken in laboratories under the same or at least similar conditions. However, these models do not perform very well on more arbitrary face images with different head poses and illumination. This thesis aims to show the challenges of Facial Expression Recognition (FER) in this wild setting. It presents the currently used datasets and the present state-of-the-art results on one of the biggest facial expression datasets currently available. The contributions of this thesis are twofold. Firstly, I analyze three famous neural network architectures and their effectiveness on the classification of facial expressions. Secondly, I present two modifications of one of these networks that lead to the proposed STN-COV model. While this model does not outperform all of the current state-of-the-art models, it does beat several ones of them.

This work covers a new approach to 3D reconstruction. In traditional 3D reconstruction one uses multiple images of the same object to calculate a 3D model by taking information gained from the differences between the images, like camera position, illumination of the images, rotation of the object and so on, to compute a point cloud representing the object. The characteristic trait shared by all these approaches is that one can almost change everything about the image, but it is not possible to change the object itself, because one needs to find correspondences between the images. To be able to use different instances of the same object, we used a 3D DPM model that can find different parts of an object in an image, thereby detecting the correspondences between the different pictures, which we then can use to calculate the 3D model. To take this theory to practise, we gave a 3D DPM model, which was trained to detect cars, pictures of different car brands, where no pair of images showed the same vehicle and used the detected correspondences and the Factorization Method to compute the 3D point cloud. This technique leads to a completely new approach in 3D reconstruction, because changing the object itself was never done before.

This thesis explores the field of artificial neural networks with realistic looking visual outputs. It aims at morphing face pictures of a specific identity to look like another individual by only modifying key features, such as eye color, while leaving identity-independent features unchanged. Prior works have covered the topic of symmetric translation between two specific domains but failed to optimize it on faces where only parts of the image may be changed. This work applies a face masking operation to the output at training time, which forces the image generator to preserve colors while altering the face, fitting it naturally inside the unmorphed surroundings. Various experiments are conducted including an ablation study on the final setting, decreasing the baseline identity switching performance from 81.7% to 75.8 % whilst improving the average χ2 color distance from 0.551 to 0.434. The provided code-based software gives users easy access to apply this neural face swap to images and videos of arbitrary crop and brings Computer Vision one step closer to replacing Computer Graphics in this specific area.

Federated Learning has recently gained more research interest. This increased attention is caused by factors including the growth of decentralized data, privacy concerns, and new privacy regulations. In Federated Learning, remote servers keep training a model on local datasets independently, and subsequently, local models are aggregated into a global model, which achieves better overall performance. Sending local model weights instead of the entire dataset is a significant advantage of Federated Learning over centralized classical machine learning algorithms. Federated learning involves uploading and downloading model parameters multiple times, so there are multiple communication rounds between the global server and remote client servers, which imposes challenges. The high number of necessary communication rounds not only increases high-cost communication overheads but is also a critical limitation for servers with low network bandwidth, which leads to latency and a higher probability of training failures caused by communication breakdowns. To mitigate these challenges, we aim to provide a fast-convergent Federated Learning training methodology that decreases the number of necessary communication rounds. We found a paper about Reducible Holdout Loss Selection (RHO-Loss) batch selection methodology, which ”selects low-noise, task-relevant, non-redundant points for training” [1]; we hypothesize, if client silos employ RHO-Loss methodology and successfully avoid training their local models on noisy and non-relevant samples, clients may offer stable and consistent updates to the global server, which could lead to faster convergence of the global model. Our contribution focuses on investigating the RHO-Loss method in a simulated federated setting for the Clothing1M dataset. We also examine its applicability to medical datasets and check its effectiveness in a simulated federated environment. Our experimental results show a promising outcome, specifically a reduction in communication rounds for the Clothing1M dataset. However, as the success of the RHO-Loss selection method depends on the availability of sufficient training data for the target RHO model and for the Irreducible RHO model, we emphasize that our contribution applies to those Federated Learning scenarios where client silos hold enough training data to successfully train and benefit from their RHO model on their local dataset.

Tools for automated segmentation and subsequent morphometric analysis of the human brain based on MRI primarily focus on healthy subjects. However, in clinical reality, this analysis is often requested for patients with various neurological conditions. Their brains might have lesions (e.g. brain tissue affected by tumors or stroke or even removed by surgery), which render many existing tools inadequate. In this study, we address this gap by utilizing an existing deep learning-based full brain segmentation network, DL+DiReCT, and retraining it with an enriched dataset of a mixed population of healthy and lesioned brains. Specifically, we increased our training set by adding annotated MRIs of two distinct groups of patients: brains after ischemic stroke and brains of patients who have undergone resective epilepsy surgery. We introduced an additional label for the lesion, aiming at enhancing the model’s lesion awareness while segmenting the rest of the brain. The modified 107-class U-Net-style segmentation model yields good overlap similarity for the new target label (i.e. DSC ≈ 0.7 for stroke lesions and DSC ≈ 0.8 for resections in epilepsy patients) while maintaining non-inferior performance compared to the baseline network on all other regions in healthy subjects. Our evaluation of the retrained model demonstrates improved robustness in brain segmentation, especially for mixed patient populations, providing essential support for volumetric estimation for clinical applications.

Plant phenotyping is the process of measuring and analyzing various traits of plants. It provides essential information on how genetic and environmental factors affect plant growth and development. Manual phenotyping is highly time-consuming; therefore, many computer vision and machine learning based methods have been proposed in the past years to perform this task automatically based on images of the plants. However, the publicly available datasets (in particular, of Arabidopsis thaliana) are limited in size and diversity, making them unsuitable to generalize to new unseen environments. In this work, we propose a complete pipeline able to automatically extract traits of interest from an image of Arabidopsis thaliana. Our method uses a minimal amount of existing annotated data from a source domain to generate a large synthetic dataset adapted to a different target domain (e.g., different backgrounds, lighting conditions, and plant layouts). In addition, unlike the source dataset, the synthetic one provides ground-truth annotations for the occluded parts of the leaves, which are relevant when measuring some characteristics of the plant, e.g., its total area. This synthetic dataset is then used to train a model to perform amodal instance segmentation of the leaves to obtain the total area, leaf count, and color of each plant. To validate our approach, we create a small dataset composed of manually annotated real images of Arabidopsis thaliana, which is used to assess the performance of the models.

Semantic image segmentation holds significance as an application in data analysis, medicine, and engineering. A primary challenge in image segmentation and object detection lies in the resource-intensive annotation process needed to create training datasets. However, the work of Bielski and Favaro [3] introduces an unsupervised approach that eliminates the need for specially annotated datasets through adversarial training. In parallel, Diffusion Denoising Probabilistic Models (DDPM) have emerged as valuable tools for synthesizing high-fidelity imagery and other media. Poole et al. [42] demonstrate the ability to repurpose DDPM as a critic to train a student network. They successfully applied this approach to train a Neural Radiance Field (NeRF) network using Score Distillation Sampling, enabling the generation of 3D representations from text prompts. In this thesis, we combine these concepts. We utilize the architecture proposed by Bielski and Favaro as a differential image parametrization and integrate the Score Distillation Sampling method from Poole et al. Our objective is to generate a layered image representation solely on a given text prompt describing a scene, without the need for ground truth data.

The use of automated systems describing the human motion has become possible in various domains. Most of the proposed systems are designed to work with people moving around in a standing position. Because such system could be interesting in a medical environment, we propose in this work a pipeline that can effectively predict human motion from people lying on beds. The proposed pipeline is tested with a data set composed of 41 participants executing 7 predefined tasks in a bed. The motion of the participants is measured with video cameras, accelerometers and pressure mat. Various experiments are carried with the information retrieved from the data set. Two approaches combining the data from the different measure technologies are explored. The performance of the different carried experiments is measured, and the proposed pipeline is composed with components providing the best results. Later on, we show that the proposed pipeline only needs to use the video cameras, which make the proposed environment easier to implement in real life situations.

Depression is the second major cause for years spent in disability and has a growing prevalence in adolescents. The recent Covid-19 pandemic has intensified the situation and limited in-person patient monitoring due to distancing measures. Recent advances in wearable devices have made it possible to record the rest/activity cycle remotely with high precision and in real-world contexts. We aim to use machine learning methods to predict an individual's mental health based on wearable-measured sleep and physical activity. Predicting an impending mental health crisis of an adolescent allows for prompt intervention, detection of depression onset or its recursion, and remote monitoring. To achieve this goal, we train three primary forecasting models; linear regression, random forest, and light gradient boosted machine (LightGBM); and two deep learning models; block recurrent neural network (block RNN) and temporal convolutional network (TCN); on Actigraph measurements to forecast mental health in terms of depression, anxiety, sleepiness, stress, sleep quality, and behavioral problems. Our models achieve a high forecasting performance, the random forest being the winner to reach an accuracy of 98% for forecasting the trait anxiety. We perform extensive experiments to evaluate the models' performance in accuracy, generalization, and feature utilization, using a naive forecaster as the baseline. Our analysis shows minimal mental health changes over two months, making the prediction task easily achievable. Due to these minimal changes in mental health, the models tend to primarily use the historical values of mental health evaluation instead of Actigraph features. At the time of this master thesis, the data acquisition step is still in progress. In future work, we plan to train the models on the complete dataset using a longer forecasting horizon to increase the level of mental health changes and perform transfer learning to compensate for the small dataset size. This interdisciplinary project demonstrates the opportunities and challenges in machine learning based prediction of mental health, paving the way toward using the same techniques to forecast other mental disorders such as internalizing disorder, Parkinson's disease, Alzheimer's disease, etc. and improving the quality of life for individuals who have some mental disorder.

The detection of interictal epileptiform discharges in the visual analysis of electroencephalography (EEG) is an important but very difficult, tedious, and time-consuming task. There have been decades of research on computer-assisted detection algorithms, most recently focused on using Convolutional Neural Networks (CNNs). In this thesis, we present the CNN Spike Detector, a convolutional neural network to detect spikes in intracranial EEG. Our dataset of 70 intracranial EEG recordings from 26 subjects with epilepsy introduces new challenges in this research field. We report cross-validation results with a mean AUC of 0.926 (+- 0.04), an area under the precision-recall curve (AUPRC) of 0.652 (+- 0.10) and 12.3 (+- 7.47) false positive epochs per minute for a sensitivity of 80%. A visual examination of false positive segments is performed to understand the model behavior leading to a relatively high false detection rate. We notice issues with the evaluation measures and highlight a major limitation of the common approach of detecting spikes using short segments, namely that the network is not capable to consider the greater context of the segment with regards to its origination. For this reason, we present the Context Model, an extension in which the CNN Spike Detector is supplied with additional information about the channel. Results show promising but limited performance improvements. This thesis provides important findings about the spike detection task for intracranial EEG and lays out promising future research directions to develop a network capable of assisting experts in real-world clinical applications.

This thesis explores the application of modern Natural Language Processing techniques to the detection of artificially generated videos of popular American politicians. Instead of focusing on detecting anomalies and artifacts in images and sounds, this thesis focuses on detecting irregularities and inconsistencies in the words themselves, opening up a new possibility to detect fake content. A novel, domain-adapted, pre-trained version of the language model BERT combined with several mechanisms to overcome severe dataset imbalances yielded the best quantitative as well as qualitative results. Additionally to the creation of the biggest publicly available dataset of English-speaking politicians consisting of 1.5 M sentences from over 1000 persons, this thesis conducts various experiments with different kinds of text classification and sequence processing algorithms applied to the political domain. Furthermore, multiple ablations to manage severe data imbalance are presented and evaluated.

The desire to use generative adversarial networks (GANs) for real-world tasks such as object segmentation or image manipulation is increasing as synthesis quality improves, which has given rise to an emerging research area called GAN inversion that focuses on exploring methods for embedding real images into the latent space of a GAN. In this work, we investigate different GAN inversion approaches using an existing generative model architecture that takes a completely unsupervised approach to object segmentation and is based on StyleGAN2. In particular, we propose and analyze algorithms for embedding real images into the different latent spaces Z, W, and W+ of StyleGAN following an optimization-based inversion approach, while also investigating a novel approach that allows fine-tuning of the generator during the inversion process. Furthermore, we investigate a hybrid and a learning-based inversion approach, where in the former we train an encoder with embeddings optimized by our best optimization-based inversion approach, and in the latter we define an autoencoder, consisting of an encoder and the generator of our generative model as a decoder, and train it to map an image into the latent space. We demonstrate the effectiveness of our methods as well as their limitations through a quantitative comparison with existing inversion methods and by conducting extensive qualitative and quantitative experiments with synthetic data as well as real images from a complex image dataset. We show that we achieve qualitatively satisfying embeddings in the W and W+ spaces with our optimization-based algorithms, that fine-tuning the generator during the inversion process leads to qualitatively better embeddings in all latent spaces studied, and that the learning-based approach also benefits from a variable generator as well as a pre-training with our hybrid approach. Furthermore, we evaluate our approaches on the object segmentation task and show that both our optimization-based and our hybrid and learning-based methods are able to generate meaningful embeddings that achieve reasonable object segmentations. Overall, our proposed methods illustrate the potential that lies in the GAN inversion and its application to real-world tasks, especially in the relaxed version of the GAN inversion where the weights of the generator are allowed to vary.

With the maturity of supervised learning technology, people gradually shift the research focus to the field of self-supervised learning. ”Momentum Contrast” (MoCo) proposes a new self-supervised learning method and raises the correct rate of self-supervised learning to a new level. Inspired by another article ”Representation Learning by Learning to Count”, if a picture is divided into four parts and passed through a neural network, it is possible to further improve the accuracy of MoCo. Different from the original MoCo, this MoCo variant (Multi-scale MoCo) does not directly pass the image through the encoder after the augmented images. Multi-scale MoCo crops and resizes the augmented images, and the obtained four parts are respectively passed through the encoder and then summed (upsampled version do not do resize to input but resize the contrastive samples). This method of images crop is not only used for queue q but also used for comparison queue k, otherwise the weights of queue k might be damaged during the moment update. This will further discussed in the experiments chapter between downsampled Multi-scale version and downsampled both Multi-scale version. Human beings also have the same principle of object recognition: when human beings see something they are familiar with, even if the object is not fully displayed, people can still guess the object itself with a high probability. Because of this, Multi-scale MoCo applies this concept to the pretext part of MoCo, hoping to obtain better feature extraction. In this thesis, there are three versions of Multi-scale MoCo, downsampled input samples version, downsampled input samples and contrast samples version and upsampled input samples version. The differences between these versions will be described in more detail later. The neural network architecture comparison includes ResNet50 , and the tested data set is STL-10. The weights obtained in pretext will be transferred to self-supervised learning, and in the process of self-supervised learning, the weights of other layers except the final linear layer are frozen without changing (these weights come from pretext).

In this thesis, we present several approaches for training a convolutional neural network using only unlabeled data. Our autonomously supervised learning algorithms are based on connections between image patch i. e. zoomed image and its original. Using the siamese architecture neural network we aim to recognize, if the image patch, which is input to the first neural network part, comes from the same image presented to the second neural network part. By applying transformations to both images, and different zoom sizes at different positions, we force the network to extract high level features using its convolutional layers. At the top of our siamese architecture, we have a simple binary classifier that measures the difference between feature maps that we extract and makes a decision. Thus, the only way that the classifier will solve the task correctly is when our convolutional layers are extracting useful representations. Those representations we can than use to solve many different tasks that are related to the data used for unsupervised training. As the main benchmark for all of our models, we used STL10 dataset, where we train a linear classifier on the top of our convolutional layers with a small amount of manually labeled images, which is a widely used benchmark for unsupervised learning tasks. We also combine our idea with recent work on the same topic, and the network called RotNet, which makes use of image rotations and therefore forces the network to learn rotation dependent features from the dataset. As a result of this combination we create a new procedure that outperforms original RotNet.

In the digital age of ever increasing data amassment and accessibility, the demand for scalable machine learning models effective at refining the new oil is unprecedented. Unsupervised representation learning methods present a promising approach to exploit this invaluable yet unlabeled digital resource at scale. However, a majority of these approaches focuses on synthetic or simplified datasets of images. What if a method could learn directly from natural Internet-scale image data? In this thesis, we propose a novel approach for unsupervised learning of object representations by mixing natural image scenes. Without any human help, our method mixes visually similar images to synthesize new realistic scenes using adversarial training. In this process the model learns to represent and understand the objects prevalent in natural image data and makes them available for downstream applications. For example, it enables the transfer of objects from one scene to another. Through qualitative experiments on complex image data we show the effectiveness of our method along with its limitations. Moreover, we benchmark our approach quantitatively against state-of-the-art works on the STL-10 dataset. Our proposed method demonstrates the potential that lies in learning representations directly from natural image data and reinforces it as a promising avenue for future research.

In computer vision, Visual Odometry is the problem of recovering the camera motion from a video. It is related to Structure from Motion, the problem of reconstructing the 3D geometry from a collection of images. Decades of research in these areas have brought successful algorithms that are used in applications like autonomous navigation, motion capture, augmented reality and others. Despite the success of these prior works in real-world environments, their robustness is highly dependent on manual calibration and the magnitude of noise present in the images in form of, e.g., non-Lambertian surfaces, dynamic motion and other forms of ambiguity. This thesis explores an alternative approach to the Visual Odometry problem via Deep Learning, that is, a specific form of machine learning with artificial neural networks. It describes and focuses on the implementation of a recent work that proposes the use of Recurrent Neural Networks to learn dependencies over time due to the sequential nature of the input. Together with a convolutional neural network that extracts motion features from the input stream, the recurrent part accumulates knowledge from the past to make camera pose estimations at each point in time. An analysis on the performance of this system is carried out on real and synthetic data. The evaluation covers several ways of training the network as well as the impact and limitations of the recurrent connection for Visual Odometry.