Hong Kong University of Science and Technology
Hong Kong University of Science and Technology Institutional RepositoryNot a member yet
162821 research outputs found
Sort by
Constrained and directional ensemble attention for facial action unit detection
No full textFacial action unit (AU) detection is a challenging task, due to the subtlety of each AU in local area and the correlations among AUs in global face. In recent years, the prevailing attention mechanism has been introduced to AU detection. However, the inherent mechanism of self-attention weight distribution has been rarely explored. Besides, ensemble learning is an efficient technique, but gains little attention in AU detection. Considering the above limitations, we propose a local self-attention constraining (LSC) network, by regarding the self-attention distribution of each AU as a spatial distribution, and constraining it based on prior knowledge so as to capture AU-related local information. Moreover, to learn correlations among different AU regions, we propose a global dual-directional attention (GDA) network, which adaptively learns global attention map from both vertical and horizontal directions. Last but not least, the two networks from different views of capturing patterns are assembled to integrate both advantages. Extensive experiments on BP4D, DISFA, and GFT benchmarks demonstrate that our methods including local self-attention constraining, global dual-directional attention, and multi-view ensemble all significantly surpass state-of-the-art AU detection works.</p
Tranexamic Acid Timing and Mortality Impact After Trauma
No full textStudy objective: Trauma resuscitation guidelines across the world have incorporated the administration of tranexamic acid (TXA) within 3 hours of injury. The 3-hour window was deduced from the Clinical Randomisation of an Antifibrinolytic in Significant Haemorrhage (CRASH-2) trial and has not been replicated. The aim of this study was to determine whether death within 28 days after trauma varied according to time from injury to the first TXA dose and, if so, precisely determine the therapeutic window. Methods: This was an exploratory analysis of the Prehospital Tranexamic Acid for Severe Trauma (PATCH-Trauma) trial, which enrolled adults with major trauma and suspected trauma-induced coagulopathy. Eligible patients were randomized to receive either TXA (administered intravenously as a bolus dose of 1 g before hospital admission, followed by a 1 g infusion over a period of 8 hours after arrival at the hospital) or matched placebo. In this analysis, we examined the effect of time from injury to first treatment dose on death within 28 days utilizing a continuous scale with linear, first-degree fractional polynomial, and second-degree fractional polynomial functions of time from injury to first treatment dose. Further log-binomial regression analyses were performed in subgroups based on the information obtained from the previous step. Results: The intention-to-treat study cohort comprised 1,287 patients, of which 635 had been allocated to the placebo arm and 652 to the TXA arm. The median time from injury to first treatment dose was 79 (interquartile range [IQR] 55 to 112) minutes. Risk of death within 28 days increased as the time to first dose of treatment increased, with benefit most pronounced up to 90 minutes. Beyond 90 minutes, the upper 95% confidence interval (CI) crossed the line of equivalence (risk ratio, 1). Administration of TXA within 90 minutes significantly reduced the risk of death within 28 days (67/393 [17%] in the TXA group versus 91/363 [25%] in placebo group; adjusted risk ratio 0.64, 95% CI 0.50 to 0.82), whereas administration beyond 90 minutes did not decrease mortality at 28 days (adjusted risk ratio 1.04, 95% CI 0.74 to 1.47). Conclusion: The optimal therapeutic window for TXA after trauma may be within 90 minutes.</p
Photoacclimation strategies of different phytoplankton species revealed by physiological, biochemical, and comparative proteomic analyses
No full textLight utilization and tolerance responses vary greatly among phytoplankton species. The molecular mechanisms underlying these differences remain elusive. Here, we investigated the growth-irradiance responses of three phytoplankton species from different functional groups. The coccolithophore Gephyrocapsa huxleyi was found to prefer lower irradiance than the diatom Phaeodactylum tricornutum and the chlorophyte Chlorella sp., which had the highest preference of the three. All three species upregulated their light-harvesting complex proteins under low light to increase light absorption and increase their cellular carbon and nitrogen fixation under elevated irradiance. G. huxleyi and P. tricornutum enhanced nitrogen assimilation and promoted nitrogen use efficiency (NUE) for carbon fixation, whereas Chlorella sp. decreased NUE under high irradiance. P. tricornutum rapidly recycles carbon skeletons at the cost of a deficiency in lipid metabolism, whereas Chlorella sp. prioritizes lipid breakdown to maintain rapid growth rates at high irradiance. By investigating the global transcriptomic expression data of these phytoplankton species, we demonstrate that different light preferences and utilization strategies largely explain their distributions and niche partitioning, which may further impact carbon and nitrogen fixation in different light regimes.</p
SurgicalGS: Dynamic 3D Gaussian Splatting for Accurate Robotic-Assisted Surgical Scene Reconstruction
No full textAccurate 3D reconstruction of dynamic surgical scenes from endoscopic video is essential for robotic-assisted surgery. While recent 3D Gaussian Splatting methods have shown promise in achieving high-quality reconstructions with fast rendering speeds, their use of inverse depth loss functions compresses depth variations. This can lead to a loss of fine geometric details, limiting their ability to capture precise 3D geometry and effectiveness in intraoperative applications. To address the limitations of existing methods, we developed SurgicalGS, a dynamic 3D Gaussian Splatting framework specifically designed for improved geometric accuracy in surgical scene reconstruction. Our approach integrates a temporally coherent multi-frame depth fusion and an adaptive motion mask for Gaussian initialisation. Besides, we represent dynamic scenes using the Flexible Deformation Model and introduce a novel normalized depth regularization loss and an unsupervised depth smoothness constraint to ensure high geometric accuracy in the reconstruction. Extensive experiments on two real surgical datasets demonstrate that SurgicalGS achieves state-of-the-art reconstruction quality, especially in precise geometry, advancing the usability of 3D Gaussian Splatting in robotic-assisted surgery. Our code is available at https://github.com/neneyork/SurgicalGS.</p
Automated “E”-aware data processing for construction ESG using building information modeling and large language model
No full textEnvironmental, Social and Governance (ESG) assessment and disclosure are critical for architecture, engineering, and construction (AEC) companies to market their financial results, reputational position, and compliance with regulatory requirements. Within this framework, the environmental (“E”) dimension presents unique and formidable data management challenges distinct from social and governance aspects. Specifically, the complex interplay of quantitative metrics and qualitative descriptions within ‘E’-aware data (e.g., measurable resource consumption alongside descriptive material sourcing practices, emissions figures coupled with compliance narratives), amplified by its sheer volume and the persistent ambiguity of environmental indicators and reporting standards, poses significant obstacles to effective ‘E’-aware data disclosure. Large Language Models (LLMs) possess inherent advantages in processing such complex environmental information due to their proficient language processing and generalization capabilities. Nonetheless, the development of LLM-based methods explicitly tailored for environmental data management within the construction sector remains underexplored. To this end, this study introduces an automated, LLM-enhanced “E”-aware data processing approach for the construction industry. The innovation of this framework is threefold. First, fifteen “E”-aware indicators are meticulously crafted to align with the specific needs of construction entities. Second, an “E”-aware algorithm, integrated within the Building Information Modeling (BIM) framework, is devised to streamline the aggregation and quantification of environmental data. Third, an LLM-enhanced complex structured data processing mechanism using retrieval augmented generation (RAG) is proposed to facilitate the efficient processing of “E”-aware data pertinent to construction projects. An illustrative case study is employed to validate the feasibility and efficacy of the proposed methodology. The results demonstrate that the developed automated RAG-LLM enhanced framework significantly advances current practice by: (1) enabling standardized “E”-aware data specifications and source mapping; (2) drastically reducing processing time for large-scale ESG documentation (saving 64.4% of time); and (3) providing a robust solution for handling multi-source, multi-format data, thereby enhancing the efficiency and reliability of environmental management and ESG disclosure in the AEC industry.</p
Converting waste into Sustainable Aviation Fuel (SAF): A systematic literature review
No full textThe global aviation sector is essential for connecting people, cultures, and economies, but it significantly contributes to greenhouse gases (GHG), exacerbating environmental concerns. This systematic literature review examines the transformation of waste into Sustainable Aviation Fuels (SAF), highlighting their potential to reduce the aviation industry's carbon footprint. The review explores waste-to-fuel technologies, such as gasification, pyrolysis, liquefaction, and Fischer-Tropsch synthesis, mainly focusing on the eight ASTM-certified bio-jet fuel production pathways, demonstrating the highest readiness levels. The study covers methodologies, case studies, and optimszation studies, identifying significant trends, advancements, and gaps in the literature to develop SAF from waste. Key findings reveal that some processes can significantly reduce CO2 emissions and improve sustainability, but challenges persist. Despite the potential of thermochemical pathways combined with oil hydro-processing and their technological readiness, the pathway's production costs remain high, and robust regulatory support is needed to scale up SAF production. Integrating pathways in a hybrid format could further offer a synergistic approach to developing SAF that combine high performance with economic and environmental sustainability. Future research should address these gaps, enhance energy and economic efficiencies, and explore innovative feedstocks and catalytic processes. The review provides valuable insights for environmentalists, industry stakeholders, engineers, and policymakers, supporting efforts to achieve sustainable aviation and global environmental goals.</p
China's model of technology leapfrog: A case study of electric vehicle policies and the development of green technology
No full textThe development of green technology is vital for driving economic growth and building low-carbon economies. In this era of technological advancement, the concept of “leapfrogging” has gained significance as emerging economies adopt advanced technologies, bypassing traditional stages of development. China, once a latecomer in the global automobile sector, strategically embraced electric vehicles (EVs) in the early 2000s and, after two decades of sustained policy and industrial efforts, has overtaken major competitors to become the world leader in the EV market. Yet there is limited systematic analysis of how China's policy design, political economy, and industrial strategies jointly enabled this leapfrogging. This study addresses this gap through a systematic policy review and a case study of China's EV sector. We analyze the drivers of government decision-making, the political economy of EV policy, the design of policy frameworks, and their impacts at both national and global levels. Our findings show how China combined early entry into a new technological field with coordinated state intervention, integrating EV technology development, manufacturing capacity, domestic demand, and industrial transformation. By tracing the evolution of EV policies over the past 20 years, the study develops a framework for understanding China's technological leapfrogging model and contributes to global debates on sustainable technology transitions by offering insights for other emerging economies to design pathways for green technology.</p
Adaptive criterion and modification of wave-particle decomposition in UGKWP method for high-speed flow simulation
No full textBenefiting from its direct modeling of physical laws in discretized space and its automatic decomposition of hydrodynamic waves and particles, the unified gas-kinetic wave-particle (UGKWP) method provides significant advantages for a wide range of multiscale physical problems, including hypersonic flow, plasma transport, and radiation transport. To achieve a more effective and efficient wave-particle decomposition in high-speed flow simulations, particularly in regions with drastic scale variations, this work investigates a scale-adaptive criterion and introduces modifications to the flux evolution of the UGKWP method. In addition to the intrinsic time-based criterion embedded in the time-dependent distribution function of UGKWP, two further criteria-based on spatial resolution and local gradients–are employed to identify the local scale and reduce the computational overhead of particles in representing near-equilibrium gas distributions. Furthermore, by aligning the evolution of hydrodynamic waves with the coefficients in the time–integration flux of the unified gas-kinetic scheme (UGKS), the modified wave representation improves consistency with particle contributions, which is especially critical when flow scales vary significantly across computational cells. The effectiveness of the adaptive UGKWP method is demonstrated through a series of benchmark cases, including hypersonic flows around a cylinder at various inflow Knudsen numbers, hypersonic flow over a slender cavity, side-jet impingement in hypersonic flow, and three-dimensional hypersonic flows over a 70∘ blunted cone with a cylindrical sting.</p
LLM-powered assistant with electrotactile feedback to assist blind and low vision people with maps and routes preview
No full textPreviewing routes to unfamiliar destinations is a crucial task for many blind and low vision (BLV) individuals to ensure safety and confidence before their journey. While prior work has primarily supported navigation during travel, less research has focused on how best to assist BLV people in previewing routes on a map. We designed a novel electrotactile system around the fingertip and the Trip Preview Assistant (TPA) to convey map elements, route conditions, and trajectories. TPA harnesses large language models (LLMs) to dynamically control and personalize electrotactile feedback, enhancing the interpretability of complex spatial map data for BLV users. In a user study with twelve BLV participants, our system demonstrated improvements in efficiency and user experience for previewing maps and routes. This work contributes to advancing the accessibility of visual map information for BLV users when previewing trips.</p
Ball-milling pretreatment-driven low-temperature waste cotton fibers-based hard carbon anodes for sodium-ion batteries
No full textBiomass-derived hard carbons represent promising low-cost anode materials for sodium-ion batteries. However, their complex preparation processes and high carbonization temperatures (>1300 °C) incur significant energy consumption and cost, hindering sustainable production. This work presents an environmentally benign strategy utilizing low-cost waste cotton fibers to produce high-performance hard carbon anodes with a low carbonization temperature of 900 °C. The innovation involves the integration of a dry ball-milling pretreatment, which facilitates the phase transformation from cellulose I to cellulose II. Comprehensive microscopic and spectroscopic analyses confirm that this phase transformation promotes the formation of a microcrystalline carbon layer with enlarged interlayer spacing during subsequent low-temperature carbonization. Moreover, the electrochemical test demonstrates the ball milling-pretreatment driven hard carbon anode delivers higher reversible capacity of 270.00 mAh g−1 compared to 166.76 mAh g−1 for the non-ball milling samples and a high initial Coulombic efficiency of 85.34 %. It also exhibits excellent rate capability and long-term cycling stability, retaining a specific capacity of 193.85 mAh g−1 under a high current density of 2 A g−1 after 1300 charge-discharge cycles. Kinetic analysis further reveals that the enhanced sodium storage stems predominantly from an insertion mechanism within the engineered microcrystalline carbon layer induced by ball milling. This work establishes a cost-effective and energy-efficient pathway for transforming textile waste into high-value anode materials, advancing the sustainable development of sodium-ion battery.</p