In recent years, strong progress has been made in creating learnable affine-equivariant models for downstream tasks such as classification. However, these models encounter increased data requirements to represent all possible transformations due to greater task complexity, while having been shown to generalize poorly to out-of-distribution data. In this work, we introduce a test-time approach for generalizing to out-of-distribution data. Namely, by utilizing a network trained to reconstruct any image that is part of a standardized training distribution, our model can infer an affine transform that moves new samples in-distribution by minimizing their reconstruction loss. With this, this approach closely matches the work of Spatial Transformer Networks, which instead learn to transform data, and inverted neural renderers for pose estimation. Through experiments, we show that this method contains a strong level of out-of-distribution translation and scale invariance, as well as a small level of rotation invariance. Namely, we show that it can handle significant transformations beyond those produced by commonly used benchmarks such as AffNIST. Using this strength we show that this method excels especially in low data regimes, outperforming existing competitors. ... Read more

Action recognition is a compelling application for event-based SNNs: by modeling time directly through neuron and synaptic dynamics, SNNs are well suited to tasks whose discriminative cues are inherently temporal. Convolutional spiking neural networks (SNNs) for event-based vision typically apply 2D spatial convolutions independently at each timestep, delegating all temporal integration to leaky integrate-and-fire (LIF) neuron dynamics. However, in fast-moving scenarios with long video lengths, this can be hard to solve by modeling temporal dependencies only with spiking neurons. We propose replacing 2D convolutions with 3D spatiotemporal convolutions in a spiking ResNet, treating 3D kernels as learned short-range temporal filters that aggregate neighboring event frames before the spiking nonlinearity. This introduces a complementary timescale into the processing hierarchy: 3D convolutions handle local frame-to-frame transitions while LIF membrane dynamics continue to integrate over longer spans, yielding two distinct levels of temporal abstraction within a single feed-forward SNN. Experiments on HARDVS and EventKitchen show that this network architecture consistently improves accuracy over a matched 2D baseline. In this work, we show that increasing the temporal kernel size yields greater accuracy gains than tuning the membrane time constant alone. Crucially, this improvement comes with no significant computational overhead, making 3D convolutions a design choice that can easily increase performance with no drawbacks.
... Read more

Event-based cameras—sensors that asynchronously record pixel-level brightness changes rather than full image frames—are often assumed to be privacy-preserving due to their sparse visual output. This thesis investigates how physical sensor configurations, including temporal bandwidth, contrast thresholds, leak noise, and spatial resolution, affect the trade-off between data utility and biometric privacy risk. We introduce a cross-domain evaluation framework that measures identity leakage using reconstructed event streams and a frozen pre-trained face recognition model. Our results show that event streams retain sufficient facial structure for accurate identification. We further identify a privacy paradox in which reducing temporal bandwidth increases attacker performance by denoising the signal, while background leak noise effectively disrupts reconstruction and lowers identification accuracy. Privacy effectiveness also varies substantially with subject motion, and the apparent benefits of resolution scaling are largely explained by domain mismatch. Overall, the findings suggest that static sensor configurations cannot guarantee anonymity, highlighting the need for threat-aware sensor design and adaptive privacy safeguards. The code used in this research is available at: [https://github.com/Stunner070/research-project-bsc-cse-tudelft](https://github.com/Stunner070/research-project-bsc-cse-tudelft). ... Read more

Event cameras are increasingly deployed in privacy-sensitive applications such as surveillance, autonomous vehicles, and human-computer interaction. Unlike conventional cameras, they record only per-pixel brightness changes as a sparse, asynchronous stream of events, making them efficient and potentially privacy-preserving. However, reconstruction models such as E2VID can recover recognizable facial images from event streams, undermining this assumption. This paper investigates whether simple perturbations applied directly to raw event streams can reduce face identifiability while preserving reconstruction quality. Three perturbation methods are compared: polarity flipping, spatial jitter, and event insertion and deletion. Each method is evaluated across multiple strength levels on 300 face video clips from the CelebV-HQ dataset, converted to synthetic events using v2e. Reconstruction quality is measured using PSNR, SSIM, and LPIPS, while face identifiability is measured using FaceNet re-identification. The implementation is available at GitHub repository. ... Read more

Facial recognition systems pose significant privacy risks, encouraging the development of generative adversarial evasion methods, such as AMT-GAN and Adv-CPG. While effective on clean, high-resolution RGB images, it remains unknown whether facial protection methods are still effective under the reconstruction pipeline of event-based cameras. This research investigates the privacy-naturalness trade-off of applying adversarial makeup to event-reconstructed faces. CelebV-HQ video clips were converted to event streams, reconstructed into grayscale images using E2VID under different thresholds, and evaluated for Attack Success Rate (ASR) and Structural Similarity (SSIM). The results reveal the following: the event-reconstruction process reduced AMT-GAN’s protection effectiveness, dropping mean ASR across four white-box models. A contrast-threshold ablation indicated this reduction is a direct result of the event-generation process itself, rather than just data loss from sparse event streams. Furthermore, a qualitative evaluation of Adv-CPG showed serious identity over-shifting and mode collapse, failing to maintain the structural diversity of the reconstructed face-image inputs. Finally, this research shows that current RGB-based adversarial protections are highly sensitive to domain shifts and fail to provide appropriate privacy for event-reconstructed vision. The scripts and jobs used in this paper can be found in the public repository: https://github.com/MateiOpr/research-project ... Read more

Event cameras are bio-inspired sensors that asynchronously measure per-pixel brightness changes, offering lower power consumption and higher temporal resolution than conventional frame cameras. These properties make them suitable for privacy-sensitive applications, such as visual localisation in AR/VR systems, where client-server architectures are used to offload computationally expensive processing from resource-limited edge devices. However, transmitting visual data to a service provider introduces privacy risks. Kim et al. propose a privacy-preserving visual localisation method that assumes an honest-but-curious service provider, but acknowledge that their approach is insufficient against a more capable attacker that can, for example, extract raw event data directly. This paper addresses this limitation by encrypting raw event camera data using the algorithm described by Zhang et al., for which no implementation was previously available. The algorithm is implemented within the visual localisation pipeline of Kim et al. and evaluated on the EvRooms dataset. The theoretical and practical effectiveness of the encryption is analysed, and improvements to the original algorithm are proposed and tested. The impact on both privacy preservation and localisation performance is measured. The paper shows that the polarity-mapping step in the implemented encryption algorithm is a powerful event data obfuscation process while still allowing retrieval of the original data. However, this process is currently not dependent on a key, which makes the algorithm not secure according to Kerckhoffs's principle. Further research should explore encryption algorithms that employ key-dependent polarity mapping. The code used in this research can be found on GitHub. ... Read more

Event-based cameras are often considered more privacy preserving than conventional RGB cameras because they don’t capture full image frames, colour, or texture. Nonetheless, their raw event streams might still encode structural information about the recorded scene. This paper questions this assumption and investigates this privacy concern experimentally by converting raw events into direct event representations and evaluating whether machine-learning models can recover semantic, spatial, and motion structure without explicit image reconstruction. Three forms of leakage are studied: semantic leakage through segmentation, spatial layout leakage through depth estimation, and motion leakage through optical flow estimation. All experiments are mainly based on DSEC dataset. As an extension dataset, PEDRo is used to investigate human-specific semantic leakage. The segmentation experiments show that semantic leakage is present but uneven: large and persistent driving scene regions such as road and background are recovered more reliably than sparse human regions in DSEC, while PEDRo shows clearer leakage of approximate human location through human-box segmentation. The depth estimation experiment shows that event representations preserve enough geometric information for a pretrained model to recover coarse scene depth. The optical flow experiment further outlines that event streams preserve recoverable motion information, since a pretrained event-based model can estimate dense motion patterns from the data. These findings highlight that event cameras don’t guarantee privacy by sensor design alone and that privacy in event-based vision depends on the representation, temporal window, task, model, and dataset. The segmentation code used in this project is available at https://github.com/ilincamaria03/event_camera_segmentation. ... Read more

Capturing dynamic processes happening inside live cells is difficult due to the fact that many subcellular structures are completely hidden beyond diffraction blur. Super-resolution optical fluctuation imaging (SOFI) surpasses the optical diffraction limit using only a conventional widefield microscope, but it requires hundreds of frames and offline cumulant computation, making it too slow for real-time live-cell imaging. Recent work (RESURF) showed that a recurrent deep network, SOFI-MISRGRU, can approximate a second-order SOFI reconstruction from as few as eight frames at real-time speed. Its fusion mechanism is inherently sequential and aggregates frames through an unnormalised summation that does not generalise beyond the training frame count. This thesis investigates whether a Transformer that fuses all frames in parallel through self-attention (TR-MISR) can improve upon this recurrent fusion in the structural fidelity, the number of frames the model can exploit, and the inference latency per output image. Quality is quantified with the resolution-scaled Pearson correlation (RSP) and a decorrelation-based resolution estimate, matching the RESURF evaluation, on a synthetic microtubule dataset spanning twelve blinking and signal-to-noise conditions. Secondly, we modify SOFI-MISRGRU into a streaming architecture that emits one super-resolved image per incoming frame. Which improved per output latency as well as allowed the model to accumulate more frames improving reconstruction fidelity. TR-MISR matches the 20-frame MISRGRU baseline while holding its quality under low SNR, slow- blinking acquisitions where MISRGRU degrades and hallucinates some filaments. Because cross- attention normalises over the input tokens, an eight-frame-trained TR-MISR can be evaluated on longer sequences without retraining, a property the GRU lacks. Our modified streaming MISRGRU reaches the batch baseline by the twentieth accumulated frame and keeps improving. Augmenting this model with a deformable phase-space alignment module and training it with a motion-augmentation pipeline that simulates moving structures further sustains reconstruction quality under inter-frame sample motion up to speeds of 5 μm per second. By emitting one image per incoming frame instead of reprocessing the whole window, the streaming model with a deformable phase-space alignment cuts the per-output latency from 170 ms to 17 ms, a 10× speed-up over the original RESURF work. Together these results bring deep-learning super-resolution microscopy closer to real-time use. ... Read more

Current inverse materials discovery methods face a trade-off between broad exploration of chemical space and control over chemical validity, synthesisability, and target properties. Here, we present the COMPosition Aware Search Strategy (COMPASS), a constrained multi-objective, multi-fidelity search framework for crystalline composition spaces. COMPASS introduces a discrete mixed-site encoding for material families with fixed site stoichiometries and up to two species mixed on each crystallographic site. This encoding preserves chemical identity, allowing empirical chemical rules and property constraints to be evaluated directly during optimisation. COMPASS combines fast composition-only screening with a constrained genetic algorithm, structure-based verification using machine-learning interatomic potentials, and active learning to improve the low-fidelity model. Applied to mixed-site ABX3 perovskites, COMPASS identifies 15,922 computationally promising candidates satisfying chemical, novelty, stability, and band-gap criteria. In the same constrained discovery task, COMPASS achieves an approximately two-orders-of-magnitude higher yield of desired candidates than the tested open-source MatterGen baselines [Zeni et al., Nature, 2025, 639, 624--632]. These results position COMPASS as a framework for chemically well-understood discovery problems where chemical constraints can guide search through large composition spaces. ... Read more

Rendering volumes using path tracing can produce stunning images, but the process is notoriously expensive. For some volumes such as clouds, there are cases where the volume has no interaction with other objects in the scene. For these cases, combined with an isotropic phase function, unidirectional path tracing from the camera computes many similar paths regardless of camera position. Computing these light paths beforehand to determine the full radiance transport through the volume speeds up the rendering stage drastically, with the result still being physically accurate.

In this thesis, we present a two-stage method designed for the aforementioned cases. The first stage is a precomputation stage, in which light paths are traced through the volume and stored in a space-efficient manner using a graph structure. Through this graph, radiance transport is then computed. In the rendering stage, the graph structure with radiance values can then be used to render the volume. The results show that our method can produce renders with negligible bias. They also show that the method is heavily constrained by space requirements as volume size increases, which results in an increase in bias for larger volumes. The main advantage of the method is its efficiency in rendering multiple images of the same object. ... Read more

The design of data collection scenarios is critical for evaluating intelligent systems for social intention recognition in aviation. Identical aircraft behaviors can generate multiple equally plausible intention interpretations depending on situational context and the observer’s professional perspective, yet existing research offers limited guidance for constructing scenarios that preserve this interpretive open-endedness. This study addresses this gap through an exploratory, literature-based investigation of how contextual factors shape intention interpretation across aviation roles. An integrated framework combining the 3Cs model of situational analysis and script theory is proposed to identify contextual dimensions influencing interpretation. Through qualitative synthesis of aviation literature, the framework demonstrates how variations in cues, classes, characteristics, and internal-external script configurations can produce divergent but valid intention narratives for the same observable behavior. The resulting scenario-first methodology provides structured guidance for designing aviation scenarios that support role-dependent intention annotation and evaluate intelligent systems. As a conceptual contribution, the framework requires empirical validation by aviation professionals. ... Read more

High-Dose-Rate brachytherapy is a critical component in the treatment of locally advanced cervical cancer. While automated treatment planning systems, such as BRIGHT, have demonstrated the ability to generate high-quality plans, their clinical adoption is hindered by the complexity of their configuration. Deploying such a system in a new hospital requires the manual definition of a clinical protocol that accurately reflects the local institution’s specific standard of care. This "cold start" problem is time-consuming for both doctors and researchers.

This thesis proposes a novel framework for the Automated Discovery of Clinical Protocols. By formulating the protocol configuration as a bi-level optimization problem, we employ the Real-Valued Gene-pool Optimal Mixing Evolutionary Algorithm to autonomously extract implicit expert knowledge from a repository of historical clinical plans. The system evolves a set of protocol parameters that, when fed into BRIGHT, reproduce radiation dose distributions as preferred by human experts.

We validate this approach using anonymized patient data from Virginia Commonwealth University. Through a series of experiments with incrementally increasing complexity, ranging from optimizing simple dose thresholds to evolving the definitions of dosimetric metrics, we demonstrate that the proposed framework can successfully identify protocols that generate treatment plans that are quantitatively similar to the clinical ground truth. This research serves as a proof-of-concept, offering a pathway to rapidly deploy automated planning systems while ensuring alignment with local clinical expertise. ... Read more

Intelligent systems are being increasingly deployed alongside humans, yet their ability to understand human intentions in order to perform tasks better is far from perfect. In this paper a literature survey will be performed to obtain existing examples of narratives, with and without intelligent systems,that take place in a Hospital - a setting with a variation in situation openness, where a framework of cues, characteristics, classes, internal scripts, external scripts, and memory and associative triggers all play an equally important role in determining an individual’s intention. Dimensions were extract from these scenarios and were analysed based on the three roles: Patients, Clinicians, and Family Members. Results and findings were used to develop a new case study involving an intelligent system assisting a doctor with diagnosing a patient and communicating it to them. Variations of contextual cues, amid other dimensions, were made to demonstrate multiple plausible narratives and its effects on intention. ... Read more

Intelligent systems in autonomous driving increasingly require the ability to infer social intentions to ensure safe and fluid interactions with human road users. However, current approaches typically frame this problem as objective trajectory prediction or fixed classification, ignoring the open-ended nature of human interpretation where a single physical behaviour can generate multiple plausible narratives. To address the gap between trajectory forecasting and narrative understanding, this research investigates how to systematically map the dimensions of variation in driving situations to the range of intention narratives they generate. A literature survey was conducted to distinguish between foundational human social norms and current algorithmic approaches. By integrating script theory with the 3Cs framework (Cues, Characteristics, Classes), this study developed a dimension extraction framework to analyse where objective observations diverge into subjective interpretations. Through comparative analysis of prototypical scenarios (lane merging and pedestrian negotiation), results revealed that current intelligent systems operate predominantly in geometric space, optimizing for physical feasibility, whereas human drivers operate in social space governed by normative scripts. The research concludes that narrative open-endedness is inversely proportional to the strength of physical and social constraints. That is, when constraints are weak, human internal scripts diverge from machine logic, leading to critical prediction errors. Consequently, future systems must shift from raw trajectory output to semantic narrative understanding to explicitly model this uncertainty and align machine reasoning with human expectations. ... Read more

Virtual Reality (VR) offers the possibility to explore and interact with complex digital worlds, yet natural locomotion is constrained by the limits of physical space. Hyperbolic geometry provides a compelling solution by embedding infinite virtual environments within finite areas, creating novel opportunities for research and design. This thesis investigates how embodied training and haptic feedback can enhance navigation and user experience in such non-Euclidean spaces. Twenty-eight participants took part in a between-subjects user study, using Holonomy VR, a hyperbolic VR application instrumented with the SenseGlove Nova 1 for force feedback and vibrotactile interaction. Participants were trained with either a drag-based embodied interface or a conventional button-based control scheme before completing matched navigation tasks. Performance was measured through speed, path efficiency and sequencing, while user experience was assessed through established questionnaires and interaction behaviour. The study finds that embodied training affords a practical advantage in subsequent navigation, and that perceived engagement with haptic elements is a strong predictor of positive usability, beyond the effects of task duration alone. Together, these results demonstrate that embodied practice and meaningful tactile interaction can help users adapt more effectively to non-Euclidean environments, offering both methodological contributions for VR research and design implications for the creation of more intuitive and engaging virtual worlds. ... Read more

Recent advances in generative AI have enabled high-quality video generation from text prompts. However, the majority of existing approaches rely exclusively on prompts, making it difficult for an artist to control the generated scene layout and motion. In this thesis, we propose a novel method for geometry-guided Text to Video generation. Our method takes as input an animated mesh sequence and a text prompt and generates a video following both the text prompt and input geometry. Our pipeline consists of two main stages: Firstly, we use an existing text-driven texture generation method to create an initial rough texture for the geometry. Next, a depth-conditioned T2I model is used to generate video frames following the guidance animation, using the generated texture to enforce temporal consistency across frames. By generating video frames rather than directly using the result of the texture generation, our method supports generating deformations from the guidance geometry and variable lighting and by using the texture for feature alignment, we acheive significantly stronger robustness to occlusions and camera motion than existing controllable video-generation approaches. We begin by identifying the failure modes of existing methods through a set of initial experiments, we then use these findings to propose our method and finally evaluate it through a series of comparisons and ablations. ... Read more

Distinguishing between benign and malignant ovarian cysts is a challenging task that depends on subjective visual markers in ultrasound scans. Current manual methods remain prone to costly misdiagnoses and the application of these methods depend heavily on the clinician's level of expertise. Recent research demonstrates promising applications of Convolutional Neural Networks (CNNs) for ovarian tumor classification; however, we observed that their performance is limited when applied to a diverse and complex dataset. To address this, we propose, implement, and evaluate three improvements to a baseline classifier.

First, we use a deep learning-based approach to remove burned-in medical annotations and introduce a weighted mean squared error (MSE) loss to improve its effectiveness by emphasizing relevant regions. This aims to better recover the original image content prior to annotation and remove annotations which can act as confounders. Second, we enhance classification by fusing image features with two readily available clinical factors at an intermediate stage of the network. Third, and central to this study, we incorporate a segmentation path that acts as a regularizer, encouraging the shared encoder to learn lesion-specific features that benefit the classification head.

These three contributions are informed by domain-specific knowledge of ovarian lesions and collectively demonstrate promising directions for improving deep learning-based models in this setting. ... Read more

Alzheimer’s Disease is a complex neurodegenerative disorder marked by the abnormal build-up of proteins in the brain. As no cure currently exists, understanding the disease’s cellular mechanisms is essential for advancing diagnostics and treatment. To this end, single-cell RNA sequencing (scRNA-seq) is a method that offers detailed information about the gene activity of individual cells but lacks their spatial context. Conversely, spatial transcriptomics technology preserves the localization of the cells but provides more limited transcriptomic information. To resolve this, we provide a model that predicts a cell’s distance to pathology from single-cell RNA-sequencing data. Additionally, we identify APOE, LYVE1, and SLC17A7 as genes potentially associated with AD-related microglial clustering around plaques. ... Read more

Single-cell RNA sequencing (scRNAseq) is a measuring technique of gene expressions in single cells that has allowed researchers to tackle Alzheimer’s disease (AD) in many ways. Single-cell data has been joined with machine learning to classify brain cells as affected by AD. However, not much is known regarding the usage of such classification models in a spatial setting. This paper analyzes how models trained on scRNAseq data can be used to find AD properties of single cells when measuring them with spatially resolved transcriptomics. With that we study the hypothesis that cells labeled as affected by the disease should appear closer to amyloid plaques, than those that are unaffected. To find out if this holds, three models are used to classify single cells spatially and their predictions are analyzed. Two single-cell datasets are used for training, each giving a drastically different classification outcome. The models do not come to a consensus on the hypothesis’ validity either, as the analysis finds no significant correlation between the variables. ... Read more

Alzheimer's Disease (AD) is a complex heterogeneous disease and is the leading cause of dementia around the world. Treatment options remain limited and the underlying mechanisms are not yet fully understood. To get more insight on this celular level, single-cell gene expression data can be used. It has proven to be effective with machine learning for tasks like cell type classification. While prior studies have explored AD classification using scRNA-seq, this has only been a binary classification. Severity of AD is classified using multiple measures, ranging from cognitive ability scores, to neuro pathological measures. This research explores the possibility of expanding the binary prediction of AD by including these measures for AD severity. In addition, given that these measures are associated, we also investigate if Multi Task Learning (MTL) models can improve the predictions by learning multiple AD related data points. If successful, this approach can give additional analysis into key tasks, genes and/or cells (sub)types that drive the models, which would lead to more possibilities for personalized treatment options, alongside more insight into the development of AD in the brain. We used a three-layer neural network architecture alongside a translation from cellular level to individual level to make individual-level predictions. Results show that Cognitive Ability can be classified best, but overal performance is only slightly above Naive Bayes. Furthermore, MTL does not appear to have any measurable positive effect on scores compared to single task models. A link to the github repository is available at \url{https://github.com/WillemDieleman/ADseverityCSE3000}. ... Read more