South African Tuberculosis Vaccine Initiative
UCT Computer Science Research Document ArchiveNot a member yet
1270 research outputs found
Sort by
Evolving Swarm-Robotic Behavioral Allocations
Full text linkThis study investigates comparative methods for two-step collective
behavior evolution (evolving group behaviors from pre-evolved
behaviors), to encourage the evolution of behavioral diversity in
swarm-robotic applications. Specifically, we investigate behavioral
diversity evolution given pre-evolved behaviors in collective behaviors
that are effective across increasingly complex and difficult
collective herding task environments. Results indicate that specific
complements of pre-evolved (lower task-performance) collective
herding behaviors was suitable for achieving high task performance
across all environments and task difficulty levels. Results support
the efficacy of the two-step approach for evolving behaviorally
heterogeneous groups in collective behavior tasks that benefit from
groups comprising various complementary behaviors
Autonomous Anomaly Detection of Orchard Tree Crown Delineations
No full textIndividual tree crown detection and delineation (ITCD) algorithms extract the boundaries of individual tree crowns from images. State-of-the-art machine learning techniques segment crowns efficiently and cost-effectively using drones and multispectral sensors. Farmers use this emerging technology to help improve crop yields, reduce resource inputs, and improve management approaches, ultimately enhancing agricultural sustainability. However, ITCD methods often fail to delineate all tree boundaries accurately. Poorly estimated delineations significantly limit the efficiency of monitoring and management strategies within orchards. This work proposes a new processing pipeline to detect three commonly occurring anomalous delineations: i) under-segmentation, ii) over-segmentation, and iii) false positives. A hand-crafted feature set comprising coarse shape descriptors, Haralick features, and spectral indices are extracted from orchard imagery to accomplish this. Furthermore, a novel approach was developed based on Geary’s c statistic in a multivariate context, leveraging physical neighbourhoods to identify contextual outlierness. This method was evaluated – using average precision and AUC-ROC performance measures – against unsupervised anomaly detection algorithms, including isolation forest (IForest), angle-based outlier detection (ABOD), and PCA-based anomaly detection. The study also trained a meta-learner to automate anomaly detection in new orchards. Results show that local outlier detection methods effectively capture over-segmentation. In contrast, global outlier detection methods better capture under-segmentation and false positives and fare well in general cases where multiple anomaly types are present. Overall, ABOD, IForest and PCA are the top models for detecting poor delineations (average AUC scores: 0.97, 0.964, and 0.962, respectively). The Geary-based approach (GBOD) gave less consistent results (average AUC score: 0.943). Nevertheless, it identified problematic regions while attributing outlierness, a capability none of the top three possess. Finally, the paper successfully demonstrates automatic model selection and anomaly detection for out-of-sample delineations, a problematic task due to the lack of reference data
Fostering co-design readiness in South Africa
No full textTechnological artifacts designed by individuals from developing nations often do not flourish in developing regions. Many authors conducting work in these regions have therefore called on researchers to make use of co-design in order to mitigate power imbalances, include the voices of communities in the design of technology intended for them, and design and build more contextually situated technology. Many of these authors also refer to the need for co-design readiness in their work. However, little is known about how co-design readiness can be articulated and achieved, especially when working with rural communities. In this paper, we make use of qualitative meta-analysis to interrogate the data from three distinct case studies in order to investigate co-design readiness. We found that co-design readiness comprises three forms, namely: (1) emotional readiness built through trust, relationship building, and empathy; (2) cultural readiness, which requires that researchers and participants have an awareness and respect for each other’s cultural values and beliefs; and, finally, (3) readiness in terms of confidence and familiarity with the technologies used which can be achieved through careful and strategic methodological decisions. We found that researchers should explicitly plan for readiness and should not assume that readiness will grow organically and that co-design, if planned well is a lengthy process. Finally, we provide learnings from our work to how co-design readiness can be achieved through planning and methodological decisions
Cross-Lingual Knowledge Projection and Knowledge Enhancement for Zero-Shot Question Answering in Low-Resource Languages
Full text linkCross-lingual knowledge projection and knowledge-enhanced language models aim to overcome the limitations of incomplete knowledge bases (KBs) and small-sized corpora in low-resource languages (LRLs). We introduce LeNS-Align, a technique that improves cross-lingual KB triple projection by combining lexical alignment, named-entity recognition, and semantic alignment. We apply LeNS-Align to project KB triples from English to four low-resource South African languages, creating more comprehensive KBs. To enhance question answering capabilities in these languages, we augment multilingual language models with Graph Neural Networks that embed the projected KB knowledge. Evaluations on three translated test sets show that our approach improves zero-shot question answering accuracy by up to 17\% compared to baselines without KB access. The results highlight how our integrated approach expands knowledge coverage and question answering capabilities in low-resource languages, addressing the challenge of scarce native KBs. This work contributes to bridging the knowledge gap for low-resource languages and demonstrates a method for enhancing NLP capabilities in resource-constrained settings
Behavior Allocations in Robotic Collective Herding Behavior Evolution
Full text linkBehavioral heterogeneity yields problem solving
benefits in biological collective behavior systems such as insect
colonies and human societies and in artificial collective behavior
systems such as distributed computer networks and swarmrobotics
systems. In this study, we investigate comparative
methods for two-step collective behavior evolution designed to
encourage the evolution of behavioral diversity in swarm robotic
applications. Specifically, we investigate behavioral diversity evolution
given pre-evolved behaviors in collective behaviors that
are effective across increasingly complex and difficult collective
herding task environments. Results indicate that a minimal complement
of pre-evolved (lower task-performance) collective herding
behaviors was suitable for achieving high task performance
across all environments and task difficulty levels. Results support
the efficacy of the two-step approach for evolving behaviorally
heterogeneous groups in collective behavior tasks that benefit
GASNet: Geometric Robust Adaptive Spatial-Enhanced Network for Building Extraction
No full textAccurately extracting building information from high-resolution remote sensing images is of great significance in urban planning, land use management, and related fields. However, challenges such as shadow interference, tree occlusion, and the complex, diverse structures of buildings make fast and accurate building extraction from remote sensing images a highly challenging task. To address this challenge, this paper proposes GASNet, a novel framework designed to enhance feature representation through global spatial dependency modeling and multiscale boundary refinement. First, we introduce the Dual-Scale Dependency Module (DSDM), which leverages graph-based reasoning in non-Euclidean space to dynamically aggregate local and global spatial dependencies while suppressing redundant features. In addition, the Scale-Aware Efficient Attention (SAEA) mechanism is proposed to enhance feature representation along both horizontal and vertical directions, enabling comprehensive boundary information capture. By leveraging the integrity of buildings, it effectively mitigates occlusion-induced interference, significantly improving the accuracy and completeness of building extraction. Extensive experiments on three benchmark datasets—WHU, Massachusetts, and Inria Aerial—demonstrate the superiority of GASNet. Our method achieves state-of-the-art performance, surpassing existing approaches by margins of 1.73% IoU on WHU, 0.48% IoU on Massachusetts, and 1.28% IoU on Inria
Using Deep Learning to Detect Stitching Artefacts in Drone-Based Orthomosaics of Orchards.
No full textThe rapidly increasing human population has necessitated the development of Precision Agriculture (PA), which leverages technology to support sustainable crop management. An important element of PA is the use of drones to capture aerial imagery of crops. In this work, we focus on orchard monitoring, specifically the image preparation phase that precedes downstream image analysis. The images produced by the drone must be stitched together to form a large orthomosaic, for later analysis. Unfortunately, the automated stitching process involves many steps and is prone to registration errors, resulting in a defective orthomosaic. Currently, these artefacts are detected through manual inspection, a time-consuming and expensive process.
We present a new automated pipeline for detecting, isolating and visualising stitching artefacts in orchard orthomosaics, built on unsupervised deep anomaly detection. To train the model, we first create orchard masks to remove non-orchard regions such as roads, houses, etc. Masked orthomosaics are then decomposed into fixed-size patches which are used to train an unsupervised deep anomaly detection model. We evaluated two anomaly detection models - reverse distillation and UniAD. During inference, a SAM-based segmentation model is used to generate an image mask, and the trained anomaly detection model assigns anomaly scores to unmasked orchard patches. For each orchard, these scores are combined to derive a final orchard classification. Two algorithms were evaluated - isolation forest and HDBSCAN clustering - over eight unique orchards.
The segmentation model provided satisfactory masking across orchards. The reverse distillation model (average AUROC 98%) outperformed the transformer-based UniAD model (89.4%). For orchard classification, HDBSCAN clustering achieved 100% accuracy with an average F1 score of 0.695 for patch classification, beating the F1 of 0.381 for isolation forest. These results were achieved using only 1200 patches for training, demonstrating the method’s potential should more data be available. Code has been made publicly available: https://github.com/dangor18/STITCH-A
Diagnostic Performance of an Artificial Intelligence Model for Detecting Pediatric Elbow Injuries on Radiographs: A Preliminary Study
No full textThis study evaluated the capability of a zero-shot AI model to detect bony and soft tissue abnormalities in pediatric elbow radiographs and assessed whether its diagnostic performance could be comparable to that of clinicians. In this retrospective cohort study, we extracted 2,700 pediatric elbow radiographs from PACS, of which 2,378 met inclusion criteria. These were split into training (1,902), validation (193), and held-out test (169) sets. The AI model (Zen-NAS with ResNet-50 backbone) was trained to detect the presence or absence of
pathology based on radiologist reports, using 13 predefined imaging categories (e.g., fractures, effusions, dislocations). Performance was assessed using sensitivity, specificity, and ROC-AUC for binary classification (pathology vs no pathology).
The AI model achieved a sensitivity of 87.19%, specificity of 94.12%, and macro-average accuracy of 81.66%. The area under the ROC curve (AUC) was 0.88, indicating strong discriminative performance. This preliminary study demonstrates that a Zen-NAS-based AI model can reliably detect pediatric elbow abnormalities on AP and lateral radiographs. While further validation is required, this technology may offer clinical value as a diagnostic support tool, particularly in settings where specialist radiology services are limited
ParentCoach: Co-designing a Chatbot to Support First-Time Parents
Full text linkRecent advancements in chatbot technology have led to their widespread application
across various sectors worldwide. Still, significant challenges remain in their e↵ective
design and implementation for healthcare in the diverse, multilingual socio-economic
contexts in South Africa. These challenges include limited internet connectivity and
the need for multilingual support.
This dissertation explores the co-design of a chatbot to support first-time parents’
informational needs in an urban South African context by drawing on the perspec-
tives of clinicians and parents using an exploratory and co-design approach. I con-
ducted one-on-one interviews with five clinicians to understand their perspectives
on parental support needs and exploratory workshops with ten parents to gather
insights on their learning challenges and experiences and their informational needs.
My analysis of findings emphasizes the importance of designing with empathy to
support vulnerable parents, ensuring chatbots complement healthcare profession-
als, building clinician trust through credible sources and endorsement by reputable
healthcare institutions, and enabling repeated access to information to aid parents’
information retention.
I then conducted two sets of co-design workshops with 21 parents that gave in-
sight into parents’ preferences regarding chatbot design modalities and uncovered
constraints for our design. These activities underscored the necessity of prepar-
ing communities to co-design unfamiliar technologies since most participants were
engaging with chatbots for the first time. Despite this unfamiliarity, participants
demonstrated an openness to adopt chatbots for parenting support.
Some key design contributions from co-design were to supplement multilingual sup-
port with English content and integrate simple language with medical terminology
to enhance parents’ understanding, enable user-initiated chatbot interactions, and
o↵er customizable features for community inclusivity.
Though we set out to co-design a chatbot to support first-time parents, I did not end
up building one due to various contextual constraints. The prototype is a “pseudo-
chatbot”, a question-and-answer informational resource presented in a chat-like user
interface with search and menus for content exploration that we evaluated in a two-
week pilot feasibility trial. The results of the trial demonstrated that familiar social
messaging interfaces and robust menu designs enhance usability, even without fully
interactive chatbot features, and highlighted the importance of aligning chatbot
content with parents’ priorities to promote engagement
Multi-Objective Evolutionary Sunshade Design
Full text linkSunshades integrated into building facade design critically influence
the building’s thermal conditions, natural lighting, energy usage,
and occupant comfort. However, heuristic designs often neglect
the multi-faceted trade-offs among these objectives. This study
compares two multi-objective evolutionary algorithms, NSGA-II
and MO-CMA-ES, in optimizing five performance metrics: thermal
comfort, Useful Daylight Illuminance (UDI), energy consumption,
outside view obstruction, and cost.We integrate annual energy and
daylight simulations, incorporating real-world weather data from
Cape Town, South Africa, and Nairobi, Kenya. Results indicate that
both MO-EAs generate Pareto-optimal sunshades exceeding the
performance of five traditional designs for all metrics. In cooler
climates, the best solutions featured upward-angled fins to admit
beneficial solar gain, while warmer climates favored configurations
blocking high-angle sunlight. These findings underscore the
importance of climate-specific optimization for identifying costeffective,
occupant-friendly building designs to balance daylight
management and energy efficiency