DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
Evaluating the printability of TI-6AL-4V and maraging steel using direct energy deposition
No full textThis project explores the printability of Ti-6Al-4V and maraging steel using direct
energy deposition process with the focus on powder characterization, chemical
composition change, defect and hardness evaluations. Both alloys are widely used in
the aerospace industry due to their excellent mechanical properties for their specific
applications. Gas-atomized powders of Ti-6Al-4V and maraging steel were first
analysed for particle size distribution, flowability, apparent density, and internal
porosity. The specimens were then printed using Makino AML900 laser metal
deposition machine under various parameter sets to assess how these factors
influenced porosity, chemical composition and hardness.
The particle size distribution analysis showed that both powders had narrow particle
size distributions and flowed well, making them suitable for the direct energy
deposition process. Chemical composition analysis revealed little change in
composition of the alloy before and after printing, suggesting little to no vaporization
during processing. Porosity levels varied depending on the process parameters, with
some builds achieving porosities below 0.4%. Hardness values were in line with
additive manufactured counterparts in literature and in some cases, were higher due to
high cooling rates of direct energy deposition.
Although detailed process optimization was beyond the scope of this work, the results
offer useful insights into the printability of Ti-6Al-4V and maraging steel using direct
energy deposition. These findings can help guide future research and development of
additively manufactured components for aerospace applications.Bachelor's degre
Religious ecology: transforming crises into opportunities for dialogue and cooperation
No full textAs the climate crisis intensifies, various religious environmental groups have emerged in a bid to encourage more people to become ecologically conscious. Such efforts have increasingly become interfaith in nature, providing an alternative space for those of different faiths to gather in dialogue and collaboration. These approaches can contribute to fostering more cohesive, resilient and compassionate societies.Published versio
Merde: Trying something new
No full textThis Final Year Project Report documents the production of the Final Year Short Documentary, titled Merde, from start to finish and from my perspective being the Director and Producer and being partnered with Joanne Leong in a two-woman team. This report will detail the intentions of the film, the specifics of the production process, working with outsiders and collaborators, and achieving the final film.Bachelor's degre
Automatically deriving developers' technical expertise from the GitHub social network
No full textDevelopers’ technical expertise is crucial for numerous tasks within open-source communities, such as identifying suitable developers and maintainers. Despite its significance, GitHub, the world’s largest open-source code hosting platform, does not explicitly display developers’ technical expertise. Existing methods fall short in capturing the multifaceted and dynamic nature of developers’ skills and knowledge. To address this gap, we propose a novel approach that leverages graph neural networks (GNNs) to express developers’ technical expertise. Our method constructs a comprehensive GitHub social network that integrates various social and development activities. We then employ a GNN model to learn a low-dimensional representation vector for each developer, encapsulating their technical expertise across different dimensions. We assess the effectiveness of our model by comparing it against five baselines on three GitHub social relationship recommendation tasks, including SimDeveloper, ContributionRepo, and RepoMaintainer. Our proposed method outperforms these baselines, achieving improvements of 5.6%-9.5% on Hit Ratio@10 and 3.4%-11.1% on F1 score. These results demonstrate promising performance in predicting technical preferences for both repositories and developers. This research contributes to a more nuanced understanding of developer expertise in open-source communities and has potential implications for improving collaboration and project management on platforms like GitHub.Submitted/Accepted versionThi work was supported by the National Key R&D Program of China (Grant No. 2023YFF0616900)
Arbitrary-steps image super-resolution via diffusion inversion
No full textThis study presents a new image super-resolution (SR) technique based on diffusion inversion, aiming at harnessing
the rich image priors encapsulated in large pre-trained
diffusion models to improve SR performance. We design
a Partial noise Prediction strategy to construct an intermediate state of the diffusion model, which serves as the
starting sampling point. Central to our approach is a
deep noise predictor to estimate the optimal noise maps
for the forward diffusion process. Once trained, this noise
predictor can be used to initialize the sampling process
partially along the diffusion trajectory, generating the desirable high-resolution result. Compared to existing approaches, our method offers a flexible and efficient sampling mechanism that supports an arbitrary number of sampling steps, ranging from one to five. Even with a single
sampling step, our method demonstrates superior or comparable performance to recent state-of-the-art approaches.
The code and model are publicly available at https:
//github.com/zsyOAOA/InvSR.Agency for Science, Technology and Research (A*STAR)Submitted/Accepted versionThis study is supported under the RIE2020 Industry Alignment Fund – Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contributions from the industry partner(s)
The confidence trap: how leaders and advisors assess the credibility of military threats
No full textWhy do political leaders perceive some military threats as credible but not others? I
present and assess the Confidence Trap Theory (CTT), which explains how leader-advisor
interactions shape credibility assessments. CTT posits that confidence – the extent to which
individuals are certain they are right – influences how leaders assimilate advice and how
advisors communicate it. These dynamics can "trap" policymakers into predictable decisionmaking
patterns, limiting the effectiveness of the advisory process. Using case studies of
President Truman and Chairman Mao’s decision-making at the start of the Korean War, I test
CTT against two competing explanations: Allies Influence Theory (AIT) and Organizational
Expertise Theory (OET). I find strong support for CTT in the U.S. case and moderate support
in the Chinese case. While both leaders made flawed assessments of each other’s military
threats, they relied on distinct advisory mechanisms. Understanding confidence dynamics can
enhance credibility assessments across diverse cultural and strategic contexts.Doctor of Philosoph
AI-empowered person re-identification based on skeleton data
No full textPerson re-identification (re-ID) plays a pivotal role in many safety-critical and daily
applications such as intelligent video surveillance and human tracking. Despite
great effort, conventional appearance-based methods using images or videos are
usually computationally expensive and vulnerable to appearance and environmental variations. With the popularization of low-cost and accurate skeleton-tracking
sensors, person re-ID via skeleton data is becoming an important emerging topic
with prominent advantages. It focuses on modeling of unique body structure and
motion patterns based on key body-joint positions, which typically enjoys smaller
data inputs and model sizes, more robust performance under shape and view variations, as well as better privacy protection without using any visual appearances.
This thesis focuses on effectively modeling and learning discriminative body structure features and motion patterns from skeleton data for person re-ID. We design
deep neural network based AI models and frameworks to perform automatic skeleton representation learning for person re-ID, so as to address the primary limitation
of most previous methods that rely on hand-crafted skeleton descriptors and often
fail to capture latent features beyond domain knowledge. In particular, we concentrate on and address several key challenges in this area. Firstly, most existing
methods require massive skeleton labels for model training, while the collection of
skeletal annotations is usually labor-expensive and unavailable in practice. Such
label dependency decreases their applicability and generality in real-world open
scenarios. Secondly, existing endeavors typically learn skeleton features from a single level (e.g., body joint level), while this intrinsically limits their ability to exploit
valuable hierarchical skeleton features and higher level patterns. It is still an open
challenge to model discriminative body structure and motion features at different
levels for person re-ID. The third challenge is exploiting latent relations between
different body joints or components to capture more recognizable patterns. The
way to model valuable body structural and actional relations from skeleton graph
representations remains to be an unresolved issue. Another crucial challenge is
devising generic auxiliary tasks or techniques (e.g., feature re-ranking) that can be
flexibly applied to different models to enhance their person re-ID performance.
To this end, this thesis explores AI-empowered solutions to specifically address the
aforementioned challenges. Firstly, to learn effective skeleton representations without labels, we propose a generic unsupervised Simple Masked Contrastive learning
(SimMC) framework for person re-ID. SimMC leverages randomly-masked skeleton subsequences to mine the most representative skeleton features (termed prototypes), and capture valuable intra-sequence skeleton semantics via contrasting motion continuity between subsequences. Empirical evaluations demonstrate effectiveness, efficiency, and generality of SimMC on multiple benchmark datasets.
To solve the second challenge of multi-level skeleton modeling, we propose a Hierarchical skeleton Meta-Prototype Contrastive learning (Hi-MPC) approach. Hi-MPC constructs coarse-to-fine body representations at different levels, and exploits them to learn most typical skeleton features (termed meta-prototypes) from multiple contrastive representation spaces, which improves the robustness and consistency of
previous skeleton prototype learning. A hard skeleton mining mechanism is devised
to learn skeleton importance and more valuable patterns. Extensive experiments
on five datasets demonstrate its state-of-the-art performance and scalability.
To address the third problem to fully capture latent body relations, we propose a Transformer-based Skeleton Graph prototype contrastive learning (TranSG)
paradigm. We devise the skeleton graph transformer to simultaneously capture
structural and actional body relations, and propose the graph prototype contrastive
learning to capture more reliable prototypes and class-related semantics. The graph
structure-trajectory prompted reconstruction is proposed to learn general high-level
skeleton semantics. Extensive experiments on five benchmark datasets demonstrate
the effectiveness of TranSG and its high generality under different scenarios.
Lastly, to achieve general model enhancement, we propose a skeleton feature reranking technique to improve different models. We devise the skeleton sequence
pooling to aggregate the most salient skeleton features. Then, both k-reciprocal
and Euclidean distance are fused to integrate neighbors’ context for more reliable
feature re-ranking. A context-based voting scheme is further proposed to better
select Rank-1 candidate for person re-ID. Empirical and qualitative evaluations on
multiple datasets show the efficacy and generality of the proposed technique.Doctor of Philosoph
Wifi-based indoor localization and human activities recognition
No full textSmart sensing based indoor human perception has been receiving increasing interest
in both academia and industry recently as a result of the promising applications
for the Internet of Things (IoT). Various sensing technologies have been applied to
enable the systems to provide higher quality services such as location-based services
and activity-based services. Among the diverse sensing technologies, WiFi sensing
has attracted more and more attention due to the ubiquitous wireless infrastructure,
non-intrusive manner, affordable expense, and privacy preservation. Received
Signal Strength (RSS) of WiFi signal indicates the overall signal power at the receiver,
which corresponds to the length of the traversed path from the transmitter.
Thus, the location of a mobile device can be estimated using RSS data from several
WiFi sensors in indoor environments. WiFi Channel State Information (CSI), describes
how the wireless signal is influenced in the surrounding environment at the
subcarrier level, where human motions will change the traversed lengths of multiple
signal paths, leading to variations in CSI data. As a result, WiFi CSI is mainly
utilized for human activity recognition (HAR). With the overwhelming development
of artificial intelligence, WiFi data is usually utilized with various learning
methods for indoor human perception, which have achieved decent performances.
However, there are still many limitations and challenges in the WiFi-based human
perception.Doctor of Philosoph
Alias-free latent diffusion models: improving fractional shift equivariance of diffusion latent space
No full textLatent Diffusion Models (LDMs) are known to have an
unstable generation process, where even small perturbations or shifts in the input noise can lead to significantly
different outputs. This hinders their applicability in applications requiring consistent results. In this work, we
redesign LDMs to enhance consistency by making them
shift-equivariant. While introducing anti-aliasing operations can partially improve shift-equivariance, significant
aliasing and inconsistency persist due to the unique challenges in LDMs, including 1) aliasing amplification during
VAE training and multiple U-Net inferences, and 2) selfattention modules that inherently lack shift-equivariance.
To address these issues, we redesign the attention modules
to be shift-equivariant and propose an equivariance loss
that effectively suppresses the frequency bandwidth of the
features in the continuous domain. The resulting alias-free
LDM (AF-LDM) achieves strong shift-equivariance and is
also robust to irregular warping. Extensive experiments
demonstrate that AF-LDM produces significantly more consistent results than vanilla LDM across various applications, including video editing and image-to-image translation. Code is available at: https://github.com/
SingleZombie/AFLDMNanyang Technological UniversityAgency for Science, Technology and Research (A*STAR)Submitted/Accepted versionThis research is supported by NTU SUG-NAP. This study is also supported under the RIE2020 Industry Alignment Fund – Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)
Reinforced continual evolution of decision-making for knowledge-driven autonomous driving
No full textSelf-driving vehicles hold the promise of transforming transportation by enhancing safety, efficiency, and accessibility. However, developing decision-making frameworks that are safe, robust, adaptive, and generalizable across diverse and complex environments remains a significant challenge. This thesis presents the design and implementation of an advanced, continually evolving autonomous driving decision-making system that addresses these critical requirements through a multi-faceted approach, spanning from simulation domain to real-world open-road experiments, paving the way for the widespread adoption of autonomous vehicles.
First and foremost, decision-making safety is ensured by integrating human priors and guidance into the reinforcement learning (RL) training process via the novel Safe Human-in-the-Loop Reinforcement Learning (Safe HiL-RL) paradigm. This approach incorporates predefined safety rules and expert-driven behaviors alongside a driving risk assessment module based on an artificial potential field (APF) to mitigate the impact of imperfect human guidance and enhance overall decision-making safety. Robustness is achieved by extensively collecting and leveraging human driving behavior data to model both aleatoric and epistemic uncertainties inherent in complex driving scenarios, thereby assisting the RL process in navigating uncertainty-intensive conditions and improving system resilience. Moreover, adaptivity is realized by incorporating human intent as part of the decision system's input, enabling autonomous driving strategies to effectively transfer to unstructured road environments such as parking lots and rural paths.
To further enhance the system's generalization capabilities, this research replaces the specialized models with foundation models pre-trained on massive, internet-scale datasets recognized for their artificial general intelligence (AGI)-level proficiencies. These foundation models provide a comprehensive suite of competencies, including object detection, scene interpretation, driving reasoning, action justification, and trajectory planning. By systematically expanding the system's domain-specific driving knowledge through a structured training paradigm, the proposed method significantly improves general knowledge and closed-loop driving performance.
Finally, the thesis explores continual evolution mechanisms to address "curse of rarity" (CoR) scenarios—safety-critical yet infrequent situations that static decision policies cannot reliably manage. Building upon the WiseAD system, an additional skill pool module is introduced, enabling selective fine-tuning of driving skills during deployment without altering the foundational model. This ensures the system can adapt to rare and critical scenarios while maintaining its generalization capabilities. By integrating the previously developed functionalities into a unified framework, the thesis establishes a continually evolving onboard decision-making system. Comprehensive evaluations across varied task domains, encompassing both simulation and real-world trials, validate the feasibility, effectiveness, and trustworthiness of the proposed methodology.
The advancements presented in this thesis have profound implications for the future of autonomous driving and its integration into society. By continuously improving the safety, reliability, and adaptability of decision-making systems, this research contributes to minimizing traffic accidents and enhancing overall road safety for passengers, pedestrians, and other road users. The incorporation of AGI-level capabilities and mechanisms for continual evolution establishes a robust foundation for the large-scale deployment of autonomous driving technologies in real-world settings. This progress not only fosters greater public trust but also accelerates the adoption of autonomous vehicles as a groundbreaking solution for intelligent transportation.
Last but not least, the methodologies developed in this work are rooted in principles of general intelligence and universal paradigms, indicating that the proposed approaches can be extended beyond the transportation sector to various other domains, such as robotics control and navigation. By leveraging these generalizable frameworks and intelligence-driven strategies, the techniques developed in this thesis can facilitate advancements in multiple fields, promoting the creation of more intelligent, adaptable, and resilient systems capable of performing complex, context-aware tasks across diverse applications.Doctor of Philosoph