Publications

Abstract

Monitoring occupancy and crowd density in indoor university spaces depends heavily on accurate person detection. However, occlusion and unfavorable camera placement often make full-body detection unreliable, and head detection offers a more practical alternative. Head detection is also a challenging small object detection problem, since heads are naturally small relative to the body and often appear overlapping, partially visible, or under uneven indoor lighting. This paper proposes Surveillant Head Detection using You Only Look Once 12 (SHD-YOLO12), a customized YOLO12 architecture for human head detection in indoor surveillance. SHD-YOLO12 integrates a Generalized Feature Pyramid Network (GFPN) style bidirectional fusion neck with Efficient Multi-Scale Attention (EMA), implemented through two custom modules, A2C2f_EMA in the backbone and C2f_EMA in the neck, to preserve fine spatial detail for small, densely packed heads. We create a surveillance head dataset from indoor university buildings and use tiling with brightness and exposure jitter to boost robustness against small and partially visible heads. SHD-YOLO12 outperforms twelve YOLO variants across nano and small models, achieving a mean Average Precision of 96.40% at IoU 0.50 (mAP@0.50) and 77.40% across IoU thresholds from 0.50 to 0.95 (mAP@0.50:0.95), representing 1.2 percentage points above the strongest baseline. SHD-YOLO12 converges faster and more stably during training than the baselines, at the cost of higher total training time, which is acceptable for surveillance deployments where detection accuracy is prioritized over training speed.

Abstract

Lung segmentation is often used to restrict chest radiograph classifiers to pulmonary anatomy, but its observed effects may vary across acquisition settings, mask sources, and model configurations. We evaluated radiology-report-derived pneumonia classification on MIMIC-CXR-JPG across two frontal views, five input conditions (raw, MedSAM3 mask/crop, and CheXMask mask/crop), and seven backbones using patient-grouped 5-fold cross-validation with a 15% outer holdout. In anteroposterior (AP) radiographs, RAD-DINO with MedSAM3 cropping achieved the numerically highest AUROC of 0.8153 (95% CI, 0.7984–0.8316). In posteroanterior (PA) radiographs, raw EVA-X-Base achieved the numerically highest AUROC of 0.8261 (95% CI, 0.8048–0.8462). Across backbones, crop-versus-raw AUROC differences were small and heterogeneous, with none remaining statistically significant after multiplicity adjustment; equivalence or noninferiority was not formally assessed. Hard masking showed larger mean AUROC reductions in PA radiographs (−0.0141 for MedSAM3 and −0.0142 for CheXMask), with four backbone-specific PA masking contrasts remaining significant after Holm adjustment. These findings do not support a universal benefit of lung-focused preprocessing within the internally validated MIMIC-CXR-JPG cohort. Cross-backbone differences were interpreted descriptively under the shared training protocol, and the generalizability of these preprocessing effects to independent institutions and imaging populations remains to be established.

Abstract

Cutting simulation is a key component of surgical simulators, requiring real-time performance, numerical stability, and physically plausible deformation. Achieving these properties is challenging because cutting combines collision detection, dynamic topology modification, and computationally intensive physical simulation. In this work, deformable objects are modeled using Position-Based Dynamics (PBD), while volumetric behavior is approximated through internal shape-preserving constraints (ISPCs) applied to triangular surface meshes. The primary contribution is a hybrid CPU–GPU cutting architecture in Unity 3D, in which the computationally expensive triangle–triangle intersection test is parallelized using compute shaders, surface rendering is performed directly from GPU-resident geometry buffers, and CPU–GPU synchronization overhead is minimized. Compared with our previous implementation, which parallelized only deformation and collision handling, the proposed framework additionally accelerates the intersection detection stage of cutting. The method was evaluated with single- and multiple-cut experiments on eight triangular meshes of increasing complexity (20–30,000 triangles). Seven of the eight models sustained interactive frame rates, and the cutting stage contributed negligibly to steady-state cost—below 0.5 ms on average even under repeated free cutting—with frame time being dominated instead by deformation and inter-part collision handling. Relative to our previous work, average frame rates improved by up to 677% over the slicing-based approach and up to 395% over the CPU-based ISPC implementation. Volume was maintained to within 0.4% of the initial volume for models of low geometric complexity. Two limitations were observed: for the high-resolution Armadillo model, mean frame cost was dominated by deformation and collision, while rare cut-finalization frames produced large worst-case spikes, keeping it below interactive rates; the Octopus model exhibited markedly reduced volume preservation, stabilizing at 73.4% of its initial volume. These results indicate that the proposed hybrid method substantially improves the efficiency of cutting simulation, but scaling to large and geometrically complex models remains future work.

Abstract

Fabric defect detection is a critical task in automated textile quality inspection, where both high localization accuracy and fast inference are required, especially for small and irregular defects. Although recent YOLO-based detectors offer a favorable balance of speed and accuracy, their convolutional architectures primarily capture local features and are less effective at detecting defects that require broader contextual understanding. Transformer-based detectors improve global context modeling but typically introduce higher computational complexity. To balance this trade-off, this study proposes a lightweight hybrid detector, YOLOv11-Transformer, which integrates multi-head self-attention (MHSA) modules into the backbone and neck of YOLOv11. The model was evaluated on a fabric defect dataset of 4375 images covering four defect categories: holes, cuts, thread errors, and metallic contamination. Comparative experiments were conducted against YOLOv5, YOLOv8, YOLOv9, YOLOv10, YOLOv11, YOLOv11-DETR, and RT-DETR under unified training conditions, with additional efficiency benchmarks against YOLOv6 and YOLOv7-Tiny. The proposed model achieved the best mAP50-95 of 94.50%, outperforming the baseline YOLOv11 by 2.40 percentage points, while maintaining efficient inference at 14.2 ms per image (approximately 70 FPS) on an NVIDIA RTX 4060 GPU. These results indicate that lightweight attention integration can improve fine-grained defect localization while preserving a practical accuracy-efficiency trade-off for automated textile inspection.

Abstract

We present a unified GPU-accelerated framework for real-time Smoothed Particle Hydrodynamics (SPH) fluid simulation with two-way rigid body coupling, secondary particle effects, and volumetric rendering, implemented entirely within the Unity game engine. The framework employs a weakly compressible SPH formulation with O(n) count sort-based spatial hashing and introduces a signed distance field (SDF) coupling system that evaluates three representative geometric primitives, sphere, cylinder, and torus, of increasing topological complexity directly on the GPU. Bidirectional force exchange is achieved through lock-free atomic compare-and-swap impulse accumulation, enabling thousands of fluid particles to interact simultaneously with each rigid body without serialization. A GPU stream compaction–based secondary particle system generates and classifies foam, spray, and bubble effects in real time, while a volumetric rendering pipeline samples fluid density into a 3D texture for SDF-composited volume rendering without surface mesh extraction. A conditional kernel dispatch strategy eliminates GPU cycles for disabled subsystems, and dynamic buffer management reduces memory pressure through runtime allocation. The system sustains above 54 frames per second at four million particles on a consumer-grade GPU, with sub-linear frame time scaling and a 1.70× speedup from dynamic buffer allocation over static pre-allocation.

Abstract

Self-collision detection has become the dominant computational bottleneck in GPU-accelerated cloth simulation, as modern parallel solvers such as XPBD have drastically reduced the cost of position updates while leaving collision resolution largely unoptimized. Existing spatial partitioning methods treat all cloth regions uniformly, saturating GPU memory bandwidth despite the fact that the vast majority of the mesh surface remains geometrically flat and collision-free at any given frame. We propose a hierarchical self-collision culling framework built upon a resolution-independent discrete curvature metric derived from the h2-normalized Laplace-Beltrami operator, integrated with a discrete Kirchhoff–Love shell model combining distance and dihedral bending constraints within XPBD. Unlike prior cache-dependent acceleration strategies, our method tightly couples curvature-driven geometric pruning with a fused GPU kernel design and shows that this stateless formulation is both faster and physically more reliable. Evaluated on meshes of 512×512 and 1024×1024 particles, our method achieves a 5.5% and 9.7% FPS improvement alongside a 34.9% and 28.4% reduction in active collision pairs, respectively, with qualitative validation via high-fidelity rendering confirming artifact-free self-contact and strict ground-plane non-penetration. Ablation results further reveal that temporal coherence, conventionally regarded as an optimization standard, strictly degrades both performance (FPS decrease of 1.4%p to 1.9%p) and physical accuracy (penetration depth increase of 36.1% to 100.0% relative to the curvature-only stage) on RTX 3070 GPU, advocating for stateless per-frame geometric evaluation as the preferred design paradigm.

Abstract

Simulating boiling phenomena in real time remains challenging due to the tightly coupled thermal physics, phase transitions, and bubble dynamics involved. This paper presents a Graphics Processing Unit (GPU)-accelerated boiling simulation framework built on Smoothed Particle Hydrodynamics (SPH) within Unity 3D. The method couples thermal conduction via the Brookshaw-style discrete Laplacian, a dual-pathway vaporization model with latent-heat removal, discrete nucleation-site control, vapor condensation, and temperature-dependent buoyancy within a single compute-shader pipeline. Bubble particles are visualized through a density-carving technique applied to the volumetric iso-surface. On a consumer-grade GPU, the framework sustains 49 frames per second (FPS) at 148 K particles and 14 FPS at 1.26 M particles, with the full boiling physics adding only 28 % overhead relative to a fluid-only baseline. Validation against the analytical Stefan problem yields less than 1.6 % error in phase-front position, and total mass is conserved to within 0.01 % across all tested configurations. A nucleation-site sweep confirms that increasing sites from 4 to 64 accelerates boiling onset from 15.5 s to 12.2 s and doubles vapor production. These results demonstrate that thermally coupled boiling can be achieved within a real-time particle-based simulation for interactive applications.

Abstract

This study discovers a real-time method for simulating progressive cutting in the Unity game engine. The proposed approach utilizes Position-Based Dynamics (PBD) to model deformable objects, making it suitable for applications such as surgical simulation training. Additionally, Unity’s compute buffers are employed to enhance computational efficiency through parallel processing. The cutting simulation operates on the surface mesh of the object, while internal deformations and volume preservation are represented using internal shape-preserving constraints (ISPCs). A series of progressive cutting experiments were conducted on various 3D models to evaluate the performance and accuracy of the algorithm. The results demonstrate that the proposed method achieves visually plausible real-time simulations of cuts.

Abstract

We propose a GPU-parallelized collision-detection and response framework for rigid-body dynamics, designed to efficiently handle densely populated 3D simulations in real time. The method combines explicit Euler time integration with a hierarchical Octree–AABB collision-detection scheme, enabling early pruning and localized refinement of contact checks. To resolve collisions, we employ a two-step response algorithm that integrates non-penetration correction and impulse-based velocity updates, stabilized through smoothing, clamping, and bias mechanisms. The framework is fully implemented within Unity3D using compute shaders and optimized GPU kernels. Experiments across multiple mesh models and increasing object counts demonstrate that the proposed hierarchical configuration significantly improves scalability and frame stability compared to conventional flat AABB methods. In particular, a two-level hierarchy achieves the best trade-off between spatial resolution and computational cost, maintaining interactive frame rates (≥30 fps) under high-density scenarios. These results suggest the practical applicability of our method to real-time simulation systems involving complex collision dynamics.