Borescope inspection is widely used in aircraft engine maintenance because it allows internal turbine components to be examined without full engine disassembly. However, the inspection process remains difficult because image-based observations provide limited geometric context for understanding where a blade defect is located and how it relates to the surrounding surface. Although 3D reconstruction offers a possible route toward more structured inspection support, turbine borescope video is challenging for standard reconstruction pipelines due to restricted viewpoints, weak texture, repetitive blade structures, reflective metallic surfaces, and unstable camera motion.

This thesis presents a defect-aware 3D inspection pipeline for turbine borescope footage. The final method combines fixed region-of-interest cropping, segmented HLOC/COLMAP-based sparse reconstruction, dense mesh generation, and reconstruction-aware transfer of precomputed image-domain defect annotations onto the recovered blade geometry. Rather than treating reconstruction as an isolated end goal, the pipeline uses reconstruction as the geometric basis for defect localization, mesh-level visualization, and inspection-oriented representation.

The final evaluated result shows that a usable defect-aware 3D representation can be produced for a selected turbine inspection sequence. The pipeline yields a sparse reconstruction, dense point cloud, Poisson mesh, defect-colored surface regions, sparse 3D defect support points, and timeline-linked defect outputs. An exploratory user evaluation did not show an overall preference for the 3D view over the current 2D image-based view, but it suggested that 3D geometry may support blade-level spatial interpretation. The current work should therefore be understood as a proof of concept that connects reconstruction, defect evidence, and visualization, while further work is needed to improve robustness, mapping confidence, and practical usability.
... Read more

Adding depth to RGB inputs (RGB-D) is known to improve model accuracy. State-of-the-art RGB-D models routinely adapt the Vision Transformer (ViT), but training ViTs purely on RGB-D is infeasible given the scarcity of depth data. A solution is using large RGB datasets to pre-train before fine-tuning on RGB-D, leveraging depth estimators to add complementary pseudo-depth to RGB datasets. We investigate the characteristics of models trained in this setup. We find that models, regardless of RGB-D fusion architecture, consistently learn simple patterns of depth utilization in the attention mechanism and across encoder layers. Our conclusions motivate the need to justify proposed depth fusion architectures against simple baselines, and to use depth fusion modules suited to the value of depth at each layer. We also show that after pre-training on pseudo-depth, fine-tuning favors pseudo- as opposed to real depth, highlighting the importance of minimizing their differences. ... Read more

Visual objects can appear in many different forms due to changes in recording conditions, including position, orientation, scale, viewpoint, and other spatial effects. Although these changes alter the image, they often do not alter the object category. In practice, robustness to geometric variation is usually obtained either by augmenting the training data with transformed examples or by introducing learned transformation modules. This makes the current models use huge data which limits their use to privileged few. This raises a fundamental question: do recognition models learn geometric transformations as reusable geometric structure, or do they learn transformed versions of specific object classes? Our experiments indicate that a significant portion of a model's learning capacity is consumed by memorizing how each category appears across the range of applied transformations. To investigate this question, we perform controlled experiments on synthetic datasets like MNIST and Fashion-MNIST. First, we quantify how much additional training data required when geometric transformations are present to reach a comparable accuracy. Second, we test whether transformation knowledge learned from one object class transfers to another. To compliment our findings on synthetic datasets we extend our results to a real world dataset namely German Traffic Sign Recognition Benchmark (GTSRB). Our results suggest that robustness to geometric transformations should not be treated as something that standard models will reliably learn from data alone. Instead, future work should focus on developing models that learn geometric transformations as reusable operations allowing current vision models to train efficiently and reduce the dependence on large amounts of training data. ... Read more

In a Convolutional Neural Network (CNN) the receptive field is a region of the input image that a kernel aggregates features from. Successive layers in a CNN expand the receptive field of a kernel by a fixed amount, and this growth corresponds to a consistent metric region across the entire input. The receptive field in a Graph Neural Network (GNN) depends on the graph topology, which changes based on the method of construction. k-Nearest Neighbor (k-NN) construction produces edges whose physical length varies with local point density to keep node degree constant, leading to non-uniform growth of the receptive field. We hypothesize that this inconsistency degrades GNN accuracy on tasks where metric distances carry meaning, and that radius-based graph construction is preferable, since fixing a distance threshold makes the receptive field grow
uniformly in metric space with each layer. We test both construction methods under the sparse, non-uniform sampling conditions typical of real-world point clouds. To enable these comparisons, we introduce PointMNIST, a 2D point cloud control dataset designed to isolate and empirically expose the fundamental differences between radius and kNN graph construction. We further show that augmenting point clouds with regularly distributed background points, a form of spatial “padding”, partially compensates for the distortions of k-NN construction, and restores the connec- tivity that radius graphs lose on sparse point clouds. ... Read more

“Don’t tune hyperparameters on the test set” is often stated in machine learning textbooks. Violating it is considered a cardinal sin that produces misleadingly optimistic results, corrupts benchmark integrity, and thus can even be interpreted as scientific fraud. Yet evidence suggests that test set hyperparameter tuning does occur in practice, making it all the more important to understand its actual consequences. So how bad is it, really? In this work we question this dogma and put it to an empirical test. We systematically study the magnitude of the performance inflation caused by tuning the hyperparameters on the test set for MNIST-1D and CIFAR-10. Our experiments show that while the effect is real and significant, it is frequently small relative to other sources of noise. In many cases, we find that tuning on the test set recovers exactly the same model as when tuning on the validation set. Most importantly, we find that the rankings of models remain preserved after tuning on the test set, and therefore that consistent test-set tuning does not invalidate benchmarks or model selection. Our results call for a more nuanced view of tuning hyperparameters on the test set, stimulating researchers to openly report test tuning. ... Read more

Fine-grained spatial control in image generation, such as segmentation-mask-to-image (S2I) that binds semantic descriptions to image regions, is a highly desirable feature when creating visual designs. Existing adapter-based S2I methods are limited to closed vocabularies and perform suboptimally on modern Multi-Modal Diffusion Transformer (MM-DiT) architectures, while open-vocabulary methods require full backbone fine-tuning, sacrificing the composability that adapters offer. We present SpaDiff, an open-vocabulary S2I adapter built on top of FLUX that does not require fine-tuning the base model. SpaDiff combines two mechanisms: (1) an adaptive attention masking strategy that binds regional prompts to their corresponding location by softly suppressing image and text tokens interactions as a learned function of denoising timestep, transformer layer index, and attention head; and (2) an mechanism for adhering to segmentation contours, explored in two variants, along with a parameter-efficient feature gating mechanism. SpaDiff achieves state-of-the-art performance among adapter-based methods on both the closed-vocabulary COCO-Stuff and open-vocabulary SACap-Eval benchmarks and approaches the performance of its non-adapter counterpart. ... Read more

Gated recurrent neural networks are commonly explained by their ability to create additive copy paths through time, which can preserve information and gradients over long sequences. This explanation is correct, but incomplete: useful gate values must themselves be learned, and this gate-learning process is also performed through recurrent computation. We study this missing learning step with controlled sequence classification tasks. We show that gated architectures do not solve long-range dependencies by architecture alone: when all training samples require long-range memory from the start, gated and non-gated recurrent models both fail. However, when training also contains short-dependency samples in which the same label relation can be learned over shorter temporal gaps, gated models can first learn a selective update behavior and then apply it to long-range samples. Diagnostic probes show that during successful training, larger gradients reach early recurrent states, and state updates depend more clearly on the input. A multi-class extension further shows that the learned behavior transfers partially beyond the subset of informative inputs that receives short-dependency samples. Overall, our results suggest that gates help not because they automatically solve long-range dependencies, but because they provide a mechanism that can be learned once the data makes the gate-learning problem simple enough. This reframes gated recurrence from an automatic solution to a learnable scaffold, and suggests that training data should be designed to expose gate-learning signals before relying on gates for long-range memory. ... Read more

Visual foundation models based on Vision Transformers often depend on large datasets and substantial computational resources, limiting their accessibility for resource-constrained research settings. This paper investigates the data efficiency of Masked Autoencoders (MAE) by studying how pre-training dataset size and mask ratio affect downstream representation quality. An MAE model is pre-trained on nested subsets of the same dataset ranging from 1k to 100k images, using different mask ratios, and then evaluated on a different downstream task dataset. The results show that MAE learns transferable representations even from small unlabeled datasets, with downstream accuracy increasing steadily as more pre-training data is used. The experiments also show that the optimal masking difficulty depends on the data regime: lower masking improves validation accuracy for the smallest subsets, while the original 75% MAE mask ratio becomes stronger as the dataset size increases. These findings suggest that mask ratio should not be treated as a fixed default in MAE training. Instead, reducing the mask ratio can improve data efficiency when pre-training data is limited, while higher masking remains effective when more visual variation is available. ... Read more

Modern computer vision often reuses a single model, trained once on many images, as a start- ing point for new tasks. Because labels are ex- pensive, a common way to train such a model is self-supervised learning (SSL), which learns from unlabeled images. SSL normally uses millions of images, and it is unclear how well it works when far fewer are available. We study one SSL method, Barlow Twins, in that case. We pre-train a small vision transformer (5.4M parameters) on parts of Tiny-ImageNet, from 1k to 100k unlabeled images, and train every run for the same 1000 epochs, so the only thing that changes is the amount of data. We then freeze each model and measure how well its features transfer to the 19 VTAB-1k tasks. Pre- training helps at every dataset size: the VTAB-1k average rises from 33.7% with 1k images to 39.2% with 100k, well above a 24.4% untrained baseline. But this average hides large differences between tasks: accuracy on natural-image tasks keeps rising with data, while accuracy on more specialized and structured tasks (medical, satellite, and geometric images) changes little. On the smallest dataset, training too long even lowers accuracy. And as the dataset grows, the checkpoint that scores best on the pre-training data moves further from the one that transfers best. At this small scale, then, the amount of data is not the only thing that matters: the kind of downstream task and the checkpoint we keep matter just as much. ... Read more

Self-supervised contrastive learning is a popular way to pre-train vision foundation models. So far, it has mostly been studied with large pre-training datasets, and it is most accessible to organizations with massive computational resources. In this work we evaluate the data-efficiency of one such method, Momentum Contrast (MoCo), and investigate how to make it work better when less data is available. We pre-train a Vision Transformer with MoCo on subsets of Tiny-ImageNet ranging from 1,000 to 100,000 images, and evaluate the learned representations on a diverse set of downstream tasks using linear probing. We investigate how the training parameters of MoCo should be chosen for a given amount of data, how the downstream accuracy scales with the amount of pre-training data, and how this scaling differs across types of downstream tasks. We find that the best parameters depend on the amount of data: the optimal number of negatives used for the contrastive objective grows with the size of the dataset, while the momentum coefficient has no single best value. We also find that pre-training is beneficial even with very little data, the downstream accuracy grows approximately log-linearly with the size of the pre-training set, and the data-efficiency growth rate is larger for tasks that are similar to the pre-training data. ... Read more

The face and its surrounding context are a strong signal for video analysis in sensitive domains, powering action recognition in forensics and longitudinal emotion analysis in medicine. However, faces are biometric data that privacy regulations such as the GDPR and HIPAA protect, forbidding their storage without protective measures. Pseudonymization solves this problem by replacing each face with a generated one, called a pseudonym. To remain useful, a pseudonymization method must satisfy three requirements: preserving the context around the face, mapping the same subject to the same pseudonym across separate videos, and avoiding any sensitive database that links subjects to their pseudonyms. No existing method satisfies all three. Face swapping preserves context but depends on a vulnerable identity database to stay consistent, while subject- and key-conditioned pseudonym generators remove that database but discard the original frame along with its context. This thesis closes the gap with SKPG-Swap: a hybrid framework in which a lightweight Subject- and Key-conditioned Pseudonym Generator (SKPG) derives a consistent pseudonym from a subject's face and a secret key, combined with a face-swap model which blends that pseudonym back into the original frame. Evaluated against bounding-box rendering strategies built on the same SKPG backbone, SKPG-Swap retains nearly all of the action-recognition accuracy of unmodified videos on UCF101 and outperforms the other pseudonymization methods on RAVDESS emotion recognition. A controlled experiment further shows that assigning a subject a consistent pseudonym identity, rather than an inconsistent one, results in more stable predictions across videos, motivating the consistency requirement. ... Read more

Convolutional neural networks (CNNs) trained on RGB images (red, green, blue channels) often exhibit sharp performance degradation under distribution shifts, as they tend to rely on superficial appearance cues such as background or texture. While depth information is known to provide complementary geometric signals that can improve robustness, most existing approaches assume access to ground-truth depth or rely on complex RGB-D architectures, limiting their applicability in practice.

In this work, we investigate whether estimated depth, obtained from a monocular RGB image, can serve as a simple and effective auxiliary signal to improve out-of-distribution (OOD) generalisation in standard CNN classifiers. Using both controlled toy experiments and real-world evaluations on the NICO++ benchmark, we compare RGB-only models against RGB-D variants that incorporate a single predicted depth channel via minimal fusion. Our results show that pseudo-depth consistently reduces OOD performance gaps across multiple CNN backbones, without degrading in-distribution accuracy. We further demonstrate that these gains persist under moderate corruption of the depth signal and disappear when geometric structure is entirely removed, indicating that the improvements stem from meaningful geometric information rather than the mere presence of an additional input channel. Furthermore, we analyse these effects through class-resolved confusion matrices and qualitative input-level examples, showing that depth specifically attenuates structured semantic confusions under domain shift.

Taken together, our findings suggest that even imperfect, predicted depth can act as a lightweight geometric inductive bias, helping CNN classifiers move away from brittle appearance-based shortcuts and toward more robust representations under domain shift.

https://gitlab.ewi.tudelft.nl/in5000/janvangemert/alexandraioana ... Read more

The circular restricted three-body problem is a canonical example of chaotic dynamics and forms the basis of many advanced spacecraft trajectory designs. This thesis investigates whether emerging artificial intelligence based generative and regression methods can reduce computational costs and enable new tools for exploring families of periodic orbits in mission design.

Generative models are evaluated for their ability to reconstruct, sample, and represent multiple periodic-orbit families and their bifurcation structure, while regression-based surrogates are assessed for unstable manifold propagation. A loss formulation that explicitly incorporates the Jacobi constant is introduced, encouraging approximate conservation of energy within the system, and penalizing in-sequence variations in Jacobi.Generative models (variational autoencoders, transformer-based diffusion models) successfully capture orbital structure and orbital family bifurcationary relationships and support efficient exploration, though differential correction is typically required to enforce physical validity. Regression-based surrogates (Kolmogorov-Arnold and deep neural networks) reproduce qualitative behaviour but remain insufficiently accurate for mission design. ... Read more

The field of computer vision research is very large and still growing. Many of these papers concern some type of inductive bias, by proposing new building blocks or alternative training methods for vision models. This type of research has enabled great progress in applications of vision models.

Computer vision concerns itself with the research and development of deep learning models that work on visual data. These vision models are already heavily integrated into society, powering real-world applications such as automated radiology in hospitals, self-driving cars, and autonomous drones. However, it takes a lot of data, in the form of datasets containing thousands or millions of images, to learn reliable vision models. This thesis explores the role that spatial biases (prior knowledge on the position and pose of objects in the image) can play in learning better and more data-efficient vision models.

We find that the practice of integrating prior knowledge on spatial biases (inductive spatial biases) can help to learn biases that are otherwise hard or impossible to learn. Though inductive bias can be difficult and time-consuming to design, and often increases inference cost, integrating inductive bias can result in better performance and greater data efficiency. This work showcases these patterns in spatial biases, specifically position bias and scale bias.

We find that position bias may be learned to some degree by models without the proper inductive bias, but that inductive bias helps to model these biases and improves performance. We show that whether learning position bias is helpful depends on the data. We contribute measures for position bias in vision models in general, as well as in Vision Transformers specifically, to enable the discovery of these findings. We propose an inductive bias on the position embedding of ViTs to better (un)learn position bias.

For scale bias, we find that existing scale-equivariant models for scale bias need to be tuned to the scale distribution of the data. We propose an inductive bias that allows scale-equivariant models to learn the scale bias of the dataset, thereby fitting the data better. We also propose an alternative parameterization of convolutions called MAGNet that can be adapted to known scale distributions present in the data. Models using MAGNets (FlexNets) can be much shallower and do not require pooling.

There are those who advocate against spending much time on inductive biases. The “bitter lesson” of Richard Sutton prescribes that we should simply add more data, not more inductive bias. However, data will run out at some point, perhaps sooner rather than later. Besides raw performance of vision models, given as much data as possible, should not be our only goal: data-deficient settings are real, plentiful, and important. Data-efficient vision models are the future of our field, and the search for appropriate inductive biases will remain an important endeavor. ... Read more

Parameter-Efficient Fine-Tuning (PEFT) methods for Transformers are designed for floating-point weights. When applied to extremely low-bit models (e.g., ternary {-1,0,1) they convert the base weights to floating point (dequantization) to add the update and then quantize again, which can diminish the benefits of aggressive quantization. We introduce a multiplicative ternary adapter that enables in-domain fine-tuning by applying an element-wise ternary mask to the base ternary Transformer weights, avoiding any dequantization to floating point and allowing direct merging back into the model. Constructed as the Kronecker product of two small trainable matrices and applied via a Hadamard product, the adapter preserves the ternary domain and merges with zero inference overhead. On a ternarized Llama-3.2-1B model, our method recovers substantial accuracy and surpasses stronger 2-bit baselines on most tasks, while retaining the efficiency advantages of ternary weights. ... Read more

Much progress in optical flow research has been driven by benchmark datasets. However, these datasets provide only limited feedback on the underlying causes of architectural failures, typically restricted to metrics such as end-point error (EPE), occlusion statistics, and large-displacement ranges. This leads to imprecise claims regarding areas consecutive models have improved upon. In this paper, we present an analysis tool that enables the generation of customisable datasets, allowing controlled variation in displacement size, camera corruptions, luminance, and other factors. We demonstrate the utility of this tool by analysing the behaviour of different architectures under varying displacement sizes and in low-light settings. ... Read more

Evacuation slides are critical aircraft safety components governed by stringent regulatory standards set by agencies like the Federal Aviation Administration and European Union Aviation Safety Agency. To comply with these standards, the maintenance and repacking of slides, currently performed manually, require operators to follow hundreds of precise steps. This process is labor-intensive, error-prone, and can result in costly delays and safety risks due to human error. Real-time visual inspection systems can help address these challenges, however, a key obstacle for real-world deployment of such systems is the scarcity of benchmark data from aerospace factory operations needed to validate and verify their performance. To enable this, we introduce the first known dataset tailored for evacuation slide inspection, comprising over 14,500 images captured under real and controlled conditions. This data aims to capture slide folding procedures of Embraer AFT evacuation slides, such that developed real-time systems can: (1) estimate the occluded position of the Pressure Relief Valve (PRV), (2) detect context-sensitive foreign objects such as packing clamps, and (3) calculate slide fold dimensions to prevent tolerance stacking errors. From this dataset, five benchmarks were constructed to evaluate performance across the three requirements. Baseline models were developed, including a PRV localization network using LSTMs, a variational autoencoder and object detection pipeline for packing clamp FOD, and a depth and reference-based slide fold measurement calculation method. When tested on our benchmarks, the depth-based measurement estimator showed precision and accuracy, the clamp FOD methodology showed high precision for images taken from specific cameras, however, the PRV position estimation remains a challenge that requires further research. Overall, our results set a foundation for the automation of visual inspection in slide packing and offer benchmarks for future research in this safety-critical inspection task. ... Read more

Human Action Recognition (HAR) models are increasingly deployed in high-stakes environments, yet their fairness across different human appearances has not been analyzed. We introduce a framework for auditing bias in HAR models using synthetic video data, generated with full control over visual identity attributes such as skin color. Unlike prior work that focuses on static images or pose estimation, our approach preserves temporal consistency, allowing us to isolate and test how changes to a single attribute affect model predictions. Through controlled interventions using the BEDLAM simulation platform, we show whether some popular HAR models exhibit statistically significant biases on the skin color even when the motion remains identical. Our results highlight how models may encode unwanted visual associations, and we provide evidence of systematic errors across groups. This work contributes a framework for auditing HAR models and supports the development of more transparent, accountable systems in light of upcoming regulatory standards. ... Read more

Test-time adaptation methods assume privileged access to model internals: parameters for fine-tuning, statistics for recalibration, or architectural components for modification. This assumption fails when models are deployed as certified systems, encrypted services, or under regulatory constraints that prohibit parameter changes. We present ITEM (Input Transformation via Entropy Minimization), the first preparation-agnostic sample adaptation method for test-time adaptation. ITEM learns input transformations that minimize prediction entropy using only gradient signals through frozen models, exploiting the principle that well-calibrated models produce low-entropy outputs on familiar data. Unlike existing sample adaptation methods that require specialized training procedures or parameter updates, ITEM works with any pre-trained model without modification or preparation requirements. Using scalar transformations as proof of concept, we demonstrate adaptation under extreme data scarcity: models trained on 10 samples per class and adapted with single calibration samples. ITEM significantly reduces performance degradation from corruption while existing methods fail or show negligible improvement. Our results establish that effective test-time adaptation is possible without model modification, architectural knowledge, or training preparation, opening new possibilities for adapting deployed models under real-world constraints. ... 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