DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
Andromeda and other stories: a short story cycle and an exegesis
No full text“Andromeda and Other Stories” is a short story cycle that includes seven short stories that are inspired by the themes of absurdity, revolt, and hope as defined by Albert Camus in his book The Myth of Sisyphus. An exegesis that details a thematic and technical breakdown of my writing decisions in the stories follow. In these stories, specific attention is paid to the role of education in the average Singaporean’s life, particularly about the beliefs and values that permeate society and shape one’s moral compass. In the recognition of absurdity, where one sees a divorce between the self and the external world, these characters work through their unease and make the choice between giving into the illusions of a world that is filled with higher purpose or choosing to live without meaning. The flash fiction form, elements of the fantastic, and disrupted chronology are literary techniques that can be found within the cycle as my attempt to explore different ways to depict life’s absurdity. The influence of short stories I have read from writers like Charles D’Ambrosio and Stuart Dybek, as well as local works from Amanda Lee Koe and Gilbert Koh will be discussed in the exegesis. Ultimately, these stories follow the journeys in which sobered individuals are coming into themselves while submerged in the routine of their narratives, similar to the average Singaporean who struggles with the stress of the everyday, in order to understand the wisdom Camus imparts in his understanding of the myth of Sisyphus.Master's degre
Addressing practical scenarios in person re-identification with semantic augmentation, normalization-diverse self-ensembles and attention-based gradient reversal
No full textThis thesis focuses on practical issues in Person Re-identification (Person ReID). Person ReID is a challenging image search problem where the goal is to identify a person from a set of candidate images across different non-overlapping cameras. Domain generalization in Person ReID is a difficult problem, as methods tailored to a single domain often struggle when tested on unseen domains. However, specialized domain generalization techniques can also underperform in simple single domain scenarios. This thesis explores domain generalization challenges in Person ReID and introduces semantic expansion methods to address these challenges. Additionally, it presents approaches capable of excelling in both intra/inter-domain scenarios, a concept we term omni-domain generalization. Our study also delves into performance imbalances across cameras in a camera surveillance network and proposes ways to mitigate these imbalances without sacrificing overall system performance. Our insights tackle the practical issues in real-world Person ReID applications.Doctor of Philosoph
Graph data query and visualization via large language models
No full textIn the evolving landscape of Large Language Models (LLMs), many applications
are being discovered daily with breakneck speed. One notable application is the
integration of LLMs into databases, enhancing the querying and visualization
processes of graph-based data, which this paper will explore. Traditional graph
databases often require complex query languages and extensive domain exper-
tise, posing a high barrier of skills required for non-specialists to fully utilize
their benefits. Specifically, this study aims to use a user-friendly interface for
natural language input, to integrate fine-tuned LLMs for coding to dynamically
generate Cypher queries for seamless interaction with graph databases. This al-
lows users to analyze and visualize interconnected data with minimal technical
expertise. By bridging the gap between technical and non-technical stakeholders,
the system simplifies decision-making processes in domains such as social net-
works, fraud detection, and supply chain optimization. Furthermore, this work
lays the foundation for future advancements in natural language interfaces for
graph databases, addressing a critical need in today’s increasingly data-driven
world.Bachelor's degre
Osteoporosis retinal prediction (ORION), predicting systemic health features from retinal images
No full textOsteoporosis is a silent but widespread condition that often goes undetected until advanced
stages. In this study, a multi-view, multimodal deep learning model was proposed to act as a
non-invasive biomarker for osteoporosis risk prediction by combining retinal fundus images with
patient clinical data. Pretrained convolutional neural networks (CNNs), such as ResNet50, were
used as feature extractors to leverage visual and non-visual cues, while a multilayer perceptron
(MLP) was used to process the patients biodata. The model was trained and evaluated on a
three-class classification task, targeting varying osteoporosis risk levels: Normal Bone Density,
Low Bone Density, and Osteoporosis. Despite a modest F1-score of 50%, the model performed
above random chance, indicating latent predictive signals in retinal imagery and patient
metadata. These findings suggest that with further optimization and larger datasets,
retinal-based screening may offer a viable early biomarker for osteoporosis risk assessmentBachelor's degre
Optimising vessel placements in anchorages with look-ahead and beam search strategies
No full textThis project tackles the issue of efficiently placing incoming vessels into anchorages as they arrive and depart. Anchored vessels take up circular areas, while anchorages can be assumed to be fixed polygonal areas. Hence, the project can be seen as a variation of the circle packing problem with a temporal dimension. This project builds on existing work by novelly introducing a look-ahead time window to maximise information available when deciding on vessel placements and complements this approach with beam search strategies to manage the runtime of the algorithm. We measure the performance of our algorithm against other benchmark algorithms in the literature using five different performance metrics. Results indicate that our algorithm outperforms competing algorithms by obtaining reductions in risk of vessel collisions of up to 23% and reductions in fuel costs of up to 22%, while losses obtained in other performance metrics are relatively small (not exceeding 16%) and statistically insignificant in some cases.Bachelor's degre
Illumination-aware character animation: a combined approach using animateanyone and IC-light
No full textThis paper proposes a novel approach to generating high-fidelity animated videos
under flexible illumination conditions by integrating two state-of-the-art models: An-
imateAnyone for image-to-video synthesis and IC-Light for advanced relighting. We
explore two primary pipelines: Animate-Then-Relight, where AnimateAnyone first
generates an animated video subsequently relit by IC-Light frame by frame, and
Relight-Then-Animate, in which the reference image is relit prior to the animation
process. Experiments conducted on a subset of the TikTok Dataset demonstrate that
both pipelines outperform the baseline Relight-A-Vid in terms of perceptual (FID) and
semantic (CLIP) metrics, with the Relight-Then-Animate approach showing particular
robustness in challenging side-light scenarios. These findings underscore the potential
of combining specialized animation and relighting modules for more realistic, con-
trollable video synthesis, paving the way for refined temporal modeling and adaptive
lighting strategies in future work.Bachelor's degre
Filial piety across sociocultural context and the life span
No full textFilial piety—children’s respect, duty, and care toward parents—is often misconceptualized despite its role in intergenerational relationships and aging societies globally. We challenge three prevalent misconceptions about filial piety: that it solely involves unwavering obedience to parents, that it exists only in Asian cultures, and that it exclusively concerns caregiving to older adult parents. Drawing from cross-cultural and developmental research, we propose an integrative framework incorporating three main dimensions (i.e., beliefs and values, affect, and behaviors) that evolve across historical time and developmental stages. This framework conceptualizes filial piety as a dynamic and multidimensional construct that varies systematically across sociocultural contexts, age groups, and historical periods. We conclude with directions for future research, specifically focusing on distinguishing dimensions of filial piety, methodological approaches for studying these developmental trajectories, and implications for understanding intergenerational relationships in context.Ministry of Education (MOE)Published versionThis work was supported by Singapore Ministry of Education Academic Research Fund Tier 1 Grants RG39/22 and RG126/23 and Yong Loo Lin School of Medicine, National University of Singapore Grant NUHSRO/2021/093/NUSMed/13/LOA (to P. Setoh)
Evaluating & integrating 3D human reconstruction methods
No full textThe project aims to evaluate recent state-of-the-arts deep learning methods in 3D human reconstruction and conceptualise ways to integrate the methods used.Bachelor's degre
Deep image colorization methods for press-on nails: overview and evaluation
No full textColorization techniques have traditionally been trained on datasets like ImageNet and
COCO, which focus on natural scenes with large-scale, diverse elements. However, the
ability of these models to generalize to domains vastly different from these datasets re-
mains underexplored. This research addresses this gap by investigating how colorization
methods perform when applied to press-on nails, a rapidly growing and highly creative
domain characterized by intricate textures, small-scale designs, and detailed embel-
lishments such as charms and 3D elements. We evaluate several colorization models,
analyzing their architectural frameworks, generalization capabilities, and performance
in meeting the unique challenges of press-on nail designs. Our goal is to assess whether
these AI-driven techniques can support nail artists by streamlining the design process
and enabling more efficient, personalized, and innovative creative workflows. Through
this study, we aim to provide insights into the potential of AI-assisted colorization to
enhance the press-on nail market.Bachelor's degre
Defending large language models against adversarial attacks using watermarking
No full textThis project investigates fingerprinting techniques for Large Language Models
(LLMs) to ensure secure model attribution while preserving performance and
efficiency. We evaluate several existing methods, including Instructional
Fingerprinting (IF), and show that they are vulnerable to adversarial unlearning attacks
which can effectively erase the embedded fingerprints. In contrast, our proposed
approach, which employs AES encryption to encode predefined trigger-response pairs,
proves robust against such attacks by preserving fingerprint integrity even after
adversarial interference. Empirical results show that fingerprinted models not only
retain their generation quality—with improved perplexity scores indicating better
fluency and coherence—but also incur minimal inference latency overhead,
confirming the harmlessness and efficiency of the method. These findings highlight
the viability of AES-based fingerprinting as a reliable and tamper-resistant mechanism
for securing LLMs, paving the way for more accountable and secure deployment in
real-world applications.Bachelor's degre