Publications

This study proposes a stateless hierarchical culling framework utilizing curvature measurements from the Laplace-Beltrami operator to resolve the self-collision bottleneck in GPU cloth simulation. 

This study demonstrates through experiments on high-resolution meshes that the proposed method reduces active collision pairs by up to 34.9% and improves FPS by up to 9.7%.

This study reveals that temporal coherence, a conventional optimization standard, actually degrades both computational performance and physical accuracy, proving the superiority of independent per-frame geometric evaluation.

This study proposes a real-time YOLO-based object detection algorithm to automatically identify the critical fifth wave (V wave) in Auditory Brainstem Response (ABR) graphs.

To create clean training data, the authors utilized a U-Net-based preprocessing technique to automatically remove existing manual annotations from clinical images.

The results demonstrate that this approach achieves high accuracy and rapid inference speeds, suggesting its strong potential as an automated support tool for diagnosing hearing disorders.

This paper presents a GPU-accelerated framework implemented in Unity3D to efficiently handle rigid-body collisions in dense, real-time simulations.

The method utilizes a hierarchical Octree-AABB structure for fast collision detection and a stabilized two-step response algorithm to resolve physical interactions accurately.

Experimental results demonstrate that this approach significantly improves scalability and maintains interactive frame rates compared to conventional collision detection methods.

This paper provides a comprehensive critical review of real-time physically-based simulation (PBS) techniques for deformable objects, filling a gap in the existing literature.

The authors categorize and analyze major methods—such as Finite Element Method (FEM), Mass-Spring Method (MSM), and Position-Based Dynamics (PBD)—based on their simulation approach and model representation.

Furthermore, the study evaluates the strengths and limitations of these techniques and discusses how deep learning is currently being used to enhance stability and real-time performance.

This paper introduces AMRG, the first end-to-end framework designed to generate narrative mammography reports using domain-specialized large vision-language models.

The authors employ a parameter-efficient fine-tuning strategy via Low-Rank Adaptation (LoRA) on the DMID dataset to establish a reproducible benchmark for this multimodal clinical task.

Experimental results demonstrate that AMRG achieves strong performance across both linguistic and clinical metrics, significantly improving diagnostic consistency and reducing hallucinations compared to existing methods.

This study investigates the capabilities of WebGPU for real-time cloth simulation, utilizing the Mass-Spring Method to overcome the computational limitations of traditional WebGL-based approaches.

Comparative evaluations reveal that the WebGPU implementation significantly outperforms WebGL, maintaining 60 frames per second even with high-resolution simulations of up to 640K nodes.

Further experiments confirm the framework's ability to handle complex real-time collisions between detailed cloth models and high-polygon 3D surfaces, demonstrating a strong balance between performance and visual realism.

This study introduces a GPU-based parallel processing framework combined with a position-based dynamics (PBD) solver to enhance real-time physics simulations for extended reality (XR) applications.

The method employs an AABB-based bounding volume hierarchy (BVH) structure for efficient collision detection and utilizes GPU-accessible 2D textures to optimize data storage and computational speed.

Experimental assessments show that this approach outperforms CPU-based simulations by up to 1705%, achieving stable real-time frame rates for complex deformable models like the Stanford Bunny.

This research presents an enhanced collision detection algorithm for real-time 3D simulations that combines Octree-based Axis-Aligned Bounding Box (AABB) spatial decomposition with the Möller method for precise triangle intersection.

The proposed framework utilizes GPU parallel processing via HLSL compute shaders in Unity3D to efficiently handle complex geometry and minimize computational load by simultaneously processing multiple computations.

Comparative evaluations demonstrate that this GPU-accelerated approach significantly improves performance, achieving collision detection speeds up to 45.62 times faster than CPU-based implementations.

This paper presents EfficientPBD, a GPU-based framework designed to optimize real-time collision detection between deformable objects and dense scene meshes in exXended Reality (XR) environments.

The method integrates view-frustum culling with a budget-based triangle selection strategy to dynamically limit collision candidates and resolves interactions using a Position-Based Dynamics solver.

Experiments on a Meta Quest 3 prototype demonstrate that the framework maintains stable interactive performance above 36 FPS with complex volumetric models, effectively preventing penetration and jitter.

This study addresses the limitations of previous surgical vision research by releasing a comprehensive public dataset and a novel framework for surgical scene segmentation.

The authors created a complex virtual surgery environment in Unity using real patient data and da Vinci instruments, which were then rendered into photorealistic images using semantic image synthesis models like SEAN and SPADE.

Evaluations with state-of-the-art segmentation models confirm that the inclusion of this high-quality synthetic data significantly improves segmentation performance in robotic surgery scenarios.