Supervised learning approaches have proven to be useful in diagnosing Osteoarthritis from X-ray images, aiding professionals in an otherwise time-consuming and subjective process. However, in the medical field, labeled data is scarce. For this reason, we investigate a contrastive self-supervised approach, SimCLR, capable of learning useful representations from unlabeled data. Specifically, we explore a core component of this method – the data augmentation techniques. While these augmentations are highly effective in introducing variability in conventional image datasets, they are too aggressive for medical images, often altering their semantic meaning. In this paper, we implement custom anatomy-aware augmentation techniques, which aim to preserve the main region of interest needed for a diagnosis. We evaluate these anatomy-aware augmentations including Gaussian blur, Contrast enhancement, Random resized crop, and Random erasing, against their classical counterparts by training multiple encoders based on different combinations of those augmentations. The findings of our study have shown that utilizing this anatomy-aware approach for all data augmentations a model uses does not lead to a significant improvement in its performance. However, selective use of anatomy-awareness on geometric-based approaches seems to show promising initial results.

Osteoarthritis is a chronic joint disease in which the protective cartilage between bones deteriorates over time, leading to pain, stiffness, and reduced mobility. Diagnosis is a time-consuming and somewhat subjective process. To address this challenge, machine learning techniques can be applied. However, training supervised models on medical images is often challenging because of the limited availability of labeled training data. Self-supervised methods, which pretrain models to learn useful features without labels, offer a potential solution to this issue. In this paper, we explore the use of Generative Adversarial Networks (GANs) as a pre-training step for osteoarthritis diagnosis. The first step is the training of a GAN on a semi-public dataset of x-ray images. In the second stage, we explore different strategies for fine-tuning the discriminator model to diagnose osteoarthritis. Our experiments suggest that while GAN-based pre-training offers slight improvements over purely supervised approaches, the performance gains remain modest.

A key advancement in model-based Reinforcement Learning (RL) stems from Transformer
based world models, which allow agents to plan effectively by learning an internal represen
tation of the environment. However, causal self-attention in Transformers can be computa
tionally redundant when most relevant information lies in only a few recent steps, making it
inefficient for environments with predominantly short-term memory dependencies. This pa
per investigates integrating alternative attention mechanisms into world models to address
these limitations. We embed inductive biases via local attention and Gaussian adaptive
attention, aiming to guide the model’s focus towards more relevant elements of the ob
servation history. We evaluate these modified architectures in four environments on the
Atari 100k benchmark under partially observable conditions. Our results show that, in
environments where relevant information is contained within a specific recent window of
observations (i.e. a short-term memory dependency), tuning local or Gaussian adaptive
attention to that window lets them significantly outperform causal attention within a lim
ited number of interactions. In Pong, the best performing Gaussian attention model raised
the mean return from–14.53 to–6.86, representing roughly a 53% improvement over the
baseline. The effectiveness of these mechanisms varies with the complexity and dynamism
of the influential variables within an environment, highlighting the importance of appro
priate prior selection and flexibility. This work highlights that leveraging influence-based
principles through inductive biases can lead to more data-efficient attention mechanisms for
world models, particularly when agents must learn from limited environment interactions
in diverse RL settings.

Planning agents have demonstrated superhuman performance in deterministic environments, such as chess and Go, by combining end-to-end reinforcement learning with powerful tree-based search algorithms. To extend such agents to stochastic or partially observable domains, Stochastic MuZero leveraged a framework that models environment uncertainty by splitting transitions into agent actions and learned stochastic outcomes. In this paper, we propose a novel architecture, FlowZero, which builds on this idea but replaces the discrete latent modeling of environment stochasticity with Conditional Normalizing Flows (CNF). This allows the model to learn a rich, continuous probability distribution over possible future states conditioned on the afterstate. The key advantage of this approach is its ability to perform exact log-likelihood evaluation, offering more precise density estimation than the evidence lower bound (ELBO) used in Stochastic MuZero. We aim to verify the proposed CNF’s capacity to overfit data and generalize to similar and larger data, and our novel agent FlowZero’s capacity to perform in a stochastic environment.

Recent advances in reinforcement learning (RL) have achieved superhuman performance in various domains but often rely on vast numbers of environment interactions, limiting their practicality in real-world scenarios. MuZero is a RL algorithm that uses Monte Carlo Tree Search with a learned dynamics model, which is trained only to predict rewards, values, and policies, without any explicit objective to match real environment transitions. This work investigates how constraining the learned model of MuZero to follow the real environment dynamics with either a temporal-consistency loss over latent states or a pixel-level observation-reconstruction loss impacts the sample efficiency of MuZero, tested under the Atari100k benchmark. We evaluate performance on Pong, Breakout, and MsPacman analyzing the impact of each loss and its sensitivity to loss weight. Our results show how the temporal-consistency loss can improve performance in certain environments while the observation-reconstruction loss fails to do so, and that both losses are highly sensitive to their weight coefficient, indicating that they might require task-based fine tuning.

This work investigates the impact of action sampling strategies on the performance of Sampled MuZero, a reinforcement learning algorithm designed for continuous control settings like robotics. In contrast to discrete domains, continuous action spaces require sampling from a proposal distribution beta during Monte Carlo Tree Search (MCTS), a process that is underexplored despite being central to the algorithm's effectiveness. We systematically study how performance is influenced by (1) the choice of beta distribution and (2) the use of progressive widening, an MCTS augmentation that samples additional actions for frequently visited search tree nodes. Our JAX-based implementation of Sampled MuZero is evaluated on the Brax HalfCheetah environment, testing beta as either a uniform distribution or the agent's policy distribution. Additionally, we examine how different progressive widening parameters affect planning depth and computational efficiency. Results show that while temperature modulation provides marginal benefits under specific conditions, progressive widening with properly calibrated parameters can improve planning depth and episode returns.

Self Supervised Learning (SSL) has been shown to effectively utilise unlabelled data for pre-training models used in down-stream medical tasks. This property of SSL enables it to use much larger datasets when compared to supervised models, which require manually labelled data. Medical classification tasks often require the identification of patterns inside a small Region Of Interest (ROI) known to be relevant for radiographic diagnosis. This contrasts standard image classification tasks, which generally rely on broader patterns. To guide a model in learning such anatomically relevant features, we investigated the hip osteoarthritis classification performance of a ROI-guided Masked Autoencoder (MAE) with a Convolutional Neural Network (CNN)-based architecture. Unlike conventional MAEs, which learn latent features by reconstructing randomly masked images, our alternative uses generated anatomical landmarks to exclusively mask the ROI or background. Contradicting similar research on Vision Transformer (ViT)-based MAEs, random masking outperformed our ROI-guided alternatives, revealing a fundamental difference in what drives performance for the two architectures, and guiding future research on more sophisticated ROI-guided masking strategies. The code is available on GitHub: https://github.com/Jasperdetweede/AnatAMAE/

Egocentric video provides a unique view into human perception and interaction, with growing relevance for augmented reality, robotics, and assistive technologies. However, rapid camera motion and complex scene dynamics pose major challenges for 3D reconstruction from this perspective. While 3D Gaussian Splatting (3DGS) has become a state-of-the-art method for efficient, high-quality novel view synthesis, variants, that focus on reconstructing dynamic scenes from monocular video are rarely evaluated on egocentric video. It remains unclear whether existing models generalize to this setting or if egocentric-specific solutions are needed. In this work, we evaluate dynamic monocular 3DGS models on egocentric and exocentric video using paired ego-exo recordings from the EgoExo4D dataset. We find that reconstruction quality is consistently lower in egocentric views. Analysis reveals that the difference in reconstruction quality, measured in peak signal-to-noise ratio, stems from the reconstruction of static, not dynamic, content. Our findings underscore current limitations and motivate the development of egocentric-specific approaches, while also highlighting the value of separately evaluating static and dynamic regions of a video.

The pandemic showed the power of video calls, as it was used to accomplish a multitude of tasks that normally required physical attendance. Expanding on the video call’s transmission of sight and sound, haptic bilateral teleoperation aims to add the sense of touch and the ability to manipulate a remote envir- onment as if the user is physically present. Haptic bilateral teleoperation is a technique where a robot mimics the actions of a human operator, who in turn receives feedback from the robot’s sensors. The technique requires low-latency haptic feedback, in the range of single-digit milliseconds, which is completely im- practical when transmitting force measurements over a large distance network. Instead, the low latency requirements can be fulfilled with Model Mediated Teleoperation (MMT) to predict the user’s perceived force feedback. With the addition of force feedback, the latency of the video feedback becomes an area of greater concern. Force and video feedback in conjunction are necessary for true immersion, but using MMT to generate real-time video feedback is im- practical. This work explores a configuration where MMT is used to generate haptic feedback and video feedback is transmitted from the remote environment to the user. Transmittance of video feedback is made possible due to its relaxed latency requirements compared to force feedback. The sources of video latency can be divided into two, latency caused by the network transmission and latency from the non-network components. While minimization of both is valuable, any time saved on the non-network components can be spent on expanding the net- work and maximizing the technology’s range. In this work, we aim to answer the question ”How can we minimize the latency caused by the non-network components of a live video feedback system?” Our approach is to design a field- programmable gate array (FPGA) configuration to minimize all latencies in a live video system excluding the camera, network, and monitor components. We demonstrate a no-network live video system with 34 ms latency, of which around 5 ms originate from the FPGA. This is a marked improvement of the 63 ms latency our department achieved using a non-FPGA solution. Utilization of cameras and monitors designed for low latencies is recommended for bet- ter results, with a hypothesized latency of 11.5 ms for a high-end non-network approach.

This thesis explores the challenge of achieving haptic bilateral teleoperation across long distances, focusing on enhancing Model Mediated Teleoperation (MMT) to support dynamic environments under high network latency. Instead of striving for perfect model alignment, we acknowledge the inevitable mismatch between the remote environment and its model at the operator. We suggest a set of design considerations and a complementary framework to make operator intent a key factor in MMT solution development. By developing a system where a user can control a remote robot arm to draw on a moving whiteboard using a novel position-altering scheme that attempts to prioritize operator intent, we apply the framework to minimize the impact of network latency on user experience. Our findings reveal that the implemented position-altering scheme significantly reduces the effects of network delays on drawing accuracy, and a user study confirms that force feedback delays are more critical to user satisfaction than visual feedback delays. The results suggest that satisfactory teleopera-tion performance can be achieved with network latency up to 50 ms, challenging the conventional 10 ms requirement. This research provides design considerations, a comprehensive framework with a position-altering policy, and an open-source software package for advancing haptic teleop-eration technologies.

In recent years the emergence of Spiking Neural Net- works (SNNs) has shown that these networks are a promis- ing alternative to traditional Artificial Neural Networks (ANNs) due to their low-power computing capabilities and noise robustness. Nevertheless, in recent approaches, they have either strayed away from spike time backpropaga- tion, used discrete time, single spike neurons, and/or limited themselves to shallow networks. These approaches limit the potential of SNNs by sacrificing significant aspects of what makes them special in order to have a functional network. It is for this reason that we believe that, the implemen- tation of residual connections, allowing a model that has multi-spiking neurons, precise time and spike time back- propagation is the path to allow these networks to truly shine. As it will solve one of this model’s most severe lim- itations which prevent it from being used to build deeper networks, the banishing of spikes at a deeper depth. Hence we aim to remedy this problem by feeding inputs from fur- ther up the network to revitalize the spike counts at deeper layers. In this paper, we explore the implementation of residual connections in precise time multi-spiking neural networks as a way of solving the disappearing spikes problem that inhibits spike time backpropagation. We explore two al- ternatives in the implementation of the residual connection and we analyze how they affect the accuracy of the network as the depth increases in both Multi-Layer Perceptrons and Convolutional Neural Networks. We also developed an ar- chitecture to allow for swapping of the fuse function in or- der to take full advantage of the flexibility that precise time provides us. Results show that the implemented residual connections allow for deeper training, which has the poten- tial to aid in the network’s performance, although in some cases some hurdles remain.

This paper presents a method that shows how a lidar-based 3D static environment can be constructed from a driving scenario and is used to aid in the creation of a digital twin from that scenario. Built with limited computational resources in mind, the resulting 3D static background reconstruction method is lightweight and nonetheless up to par with similar works, proving itself to be a viable alternative to similar methods. It uses ground truth labels and open-source, out-of-the-box working building blocks to create the pipeline that filters the lidar frames, performs registration between those frames, and meshes the resulting point cloud into a 3D mesh of the static background. We performed experiments in the Siemens Prescan ADAS Simulator of a reenactment of the traffic scenario in combination with our 3D static background and measured the inference domain gap using the Bidirectional Chamfer Distance and a real-data trained object detector, comparing both the synthetic and original data. Using Prescan as the simulator, AD models van be evaluated closed-loop. This work opens up the possibilities of applying sensor domain shifts to existing data or creating reality-based traffic scenarios from existing traffic scenarios and their corresponding 3D static background of edge cases that do not, or too little, exist in real-world data.

Structure-from-Motion (SfM) and Neural Radiance Fields (NeRFs) have significantly advanced 3D reconstruction in multi-view scenarios. Despite their success in handling non-repetitive, texture-rich scenes, applying such techniques to real-world scenarios with texture-less and repetitive structures remains challenging. One such case is in industrial inspection processes, specifically the reconstruction of aircraft engine blades. The goal is to build a 3D model of all the engine blades from a single video. We explore the application of SfM in modeling repetitive and texture-less blades, identify common causes of failure, and improve upon the default SfM by replacing the commonly used exhaustive match with sequential match to handle ambiguity stemming from repetitiveness. Sequential matching enables more precise pose estimation and better 3D reconstruction in our scenario. In addition, we explore the importance of choosing the correct camera model and provide a comparative look at the existing 3D mesh reconstruction solutions, presenting tweaked versions that result in better performance. This work lays the foundation for 3D reconstruction of repetitive and texture-less objects by proposing sequential matching, enabling better 3D modeling of engine blades compared to classic SfM pipelines.

Interactive video games often use vision-based systems or wearables to track player movements. Vision-based systems are privacy-invasive, and wearables require frequent recalibration and recharging. Frequently-Modulated Continuous-Wave (FMCW) radars have been proposed as an alternative tracking solution addressing these problems. Working in the millimeter-wave (mmWave) range, they capture scenes as point clusters, ensuring privacy without attaching sensors to the user. Previous research has shown their applicability in rehabilitation, gait recognition, and smart home appliances. This study focuses on integrating an mmWave sensing device with an interactive version of the Breakout video game. We propose a general framework with three main modules - generating points, clustering them, and reconstructing the player's movements as in-game commands. Our system enhances an already existing Kalman filter-based tracking algorithm. Online experiments were conducted to compare the proposed system to the baseline algorithm. The proposed system decreases the standard deviation on the estimated target location by 33% against motionless targets, while maintaining the baseline accuracy when tested on moving targets. Furthermore, it allows a higher game refresh rate, thus smoothing in-game movements. These results demonstrate the potential of FMCW radars in enhancing interactive video game experiences.

Human Pose Estimation using Millimeter Wave radars has emerged as a promising alternative to traditional camera-based systems, addressing privacy and deployment constraints. While state-of-the-art Deep Learning models predominantly focus on spatial feature extraction to determine the positions of key points in the human body, this research investigates the effects of incorporating temporal dynamics in such models. It focuses of modifying an existing state-of-the-art spatial model to account for temporal dynamics and compares the performance of the two models. Long Short-Term Memory networks are used to capture temporal dependencies between frames of point clouds which significantly boosts the precision of key point detection. The proposed temporal model demonstrates a 53% reduction in Mean Absolute Error and a 45% reduction in Root Mean Squared Error compared to state-of-the-art model. Moreover, these improvements were achieved with a less complex model architecture and similar training times. The robustness of the model was further validated on a different dataset, showcasing its potential for broad application in fields such as healthcare, sports analysis, traffic monitoring and robotics. This study underscores the efficacy of temporal dynamics in pose estimation, and showcases the advantages of accounting for temporal dependencies when evaluating more complex movements.

The People Counting Problem requires calculating the number of people in a region of interest. This is needed in crowd-monitoring scenarios but has become increasingly problematic when relying on video cameras, as they raise privacy concerns. Instead, we propose using a mmWave radar to detect people by creating point clouds from their radar signal reflections. This approach, however, can pose challenges when people walk closely together because their individual point clouds overlap and are seen as a single, larger cloud. It is difficult to count how many individuals this large point cloud holds, which can lead to miscounting the people in the scene. One approach to address this issue is leveraging the time dimension in people walking sequences, which can be done with Long Short-Term Memory (LSTM) models. Given this, we investigate how two state-of-the-art models, PointNet and MARS, perform for people counting from point clouds when extended through LSTMs. The results show how both PointNet and MARS improve performance when extended by LSTMs. Particularly, despite having over double the parameters, MARS+LSTM outperforms PointNet+LSTM in terms of accuracy and computational efficiency. MARS+LSTM can effectively capture small changes in the local structure of point clouds between frames, which PointNet loses due to max pooling. This highlights the importance of selecting a model architecture, like the CNN in MARS, that aligns with the data characteristics to maximise performance.

Crowd-control is an emerging problem in urban areas and cameras are commonly seen as the solution, however, there are major concerns regarding privacy. To overcome these issues, while still maintaining the ability to keep track of people, mmWave sensors can be utilized instead, but this introduces new challenges when it comes to people counting. They work by recording the data as point clouds, making it difficult to determine the real number of people when they are occluded. To address this challenge, we combine the spatial and temporal information of the point clouds into graphs, and explore the possibilities of Graph Neural Networks. Our system classifies up to 5 people and bikes with an accuracy of 80.47%. This accuracy is 2.53% worse than the state of the art people counting system for mmWave radar point clouds.

The modern workplace often exposes individuals to privacy risks, such as the unauthorised visibility of their computer screens. MARS (mmWave-based Assistive Rehabilitation System for Smart Healthcare), coupled with VideowindoW screens, offers an innovative solution to these threats by using mmWave radar to reconstruct human poses and estimate the position of 19 key joints. This enables the screens to become opaque based on the viewer's position, ensuring privacy. Although originally designed as a rehabilitation system, MARS can be utilised for its pose estimation capabilities to enhance workplace privacy. This research explores two modifications to the MARS architecture to assess their impact on the system's accuracy and performance. Specifically, we modify the MARS architecture by increasing the depth of its convolutional neural network (CNN) and integrating the PointNet architecture. Results establish that an optimal CNN configuration with two convolutional and two dense layers, followed by the output layer, modestly improves joint location estimation. However, integrating PointNet does not improve performance, likely due to PointNet's limitations in capturing the necessary local structural details of point clouds. These findings inform future research of possible improvements when leveraging the MARS dataset in the fields of privacy enhancement and smart healthcare applications.

Gaussian Splatting is a successful recent method for generating novel views of a scene based on photographs taken from that scene [1]. It uses rasterization in order to render the scenes it generates, which consist of 3D Gaussians. However, modern hardware and tools are designed and optimized around rendering polygonal and texture based models [2]. This paper proposes a method of extracting both a 3D model and texture file from a Gaussian Splatting scene by using renders of that scene in Photogrammetry. It shows that this can be a viable method for generating a traditional 3D model from Gaussian Splatting scene, and can for certain cases generate a model of comparable quality while lowering the number of initial images required by up to three times.

3D Gaussian Splatting (3DGS) is a method for representing 3D scenes, but is prone to overfitting when trained with limited viewpoint diversity, of- ten resulting in artifacts like floating Gaussians at incorrect depths. This paper addresses this issue by introducing 3D Gaussian Splatting with Depth, which incorporates depth supervision from RGB Depth (RGB-D) cameras into the training process. By using depth data to guide the placement of Gaussians, the proposed method aims to reduce artifacts. Through quantitative and qualitative analysis, this paper demonstrates that depth-supervised Gaussian splatting mitigates overfitting artifacts, particularly in outdoor scenes with a mediocre cam- era point diversity. The depth-supervised model is able to reduce the depth loss by a factor of three times without substantially increasing the loss on regular views.