Selected work

Representative Papers

Computer Vision-Based Early Detection of Container Loss at Sea

Apr 27, 2026

This study addresses the critical risk of container stack collapse and loss overboard during vessel motion under harsh sea conditions, which poses significant safety, environmental, and economic threats. To mitigate this challenge, the authors propose a low-cost computer vision system leveraging existing onboard cameras—requiring no additional sensors—that enables container-level micro-motion detection and quantification of relative displacement for the first time. By integrating object segmentation, optical flow-based temporal tracking, and residual motion analysis, the method effectively isolates and monitors inter-tier movements within container stacks in real-world shipborne video footage. The approach significantly enhances cargo safety, operational resilience, and compliance with International Maritime Organization (IMO) regulations, providing a foundation for early-warning systems and timely intervention.

0 citationsRead paper

Eyes on Target: Gaze-Aware Object Detection in Egocentric Video

Nov 03, 2025

To address the neglect of human visual attention in first-person video object detection, this paper proposes a gaze-guided Vision Transformer framework. It models eye-tracking trajectories as dynamic attention priors and injects them into the multi-head self-attention mechanism; a gaze-aware attention head importance metric is further introduced to explicitly modulate each head’s response strength to fixated regions. Additionally, depth information is fused to enhance spatial perception. The method achieves significant improvements over non-gaze-aware baselines—up to +3.2–5.7% mAP—on a custom simulator dataset and public benchmarks including Ego4D Ego-Motion and Ego-CH-Gaze. Ablation studies validate the effectiveness of both the gaze-guidance mechanism and the depth-fusion module. This work is the first to systematically demonstrate the interpretability value of eye movement signals for evaluating first-person object detection, establishing a novel paradigm for attention modeling in embodied intelligence.

0 citationsRead paper

ResNet: Enabling Deep Convolutional Neural Networks through Residual Learning

Oct 27, 2025

To address the training difficulty of ultra-deep convolutional neural networks (CNNs) caused by vanishing gradients, this paper proposes a residual learning framework. It introduces identity-mapping shortcut connections that allow input signals to bypass certain nonlinear transformations, thereby ensuring unimpeded forward signal propagation and backward gradient flow. This approach enables stable end-to-end training of CNNs with over 100 layers—e.g., ResNet-110—on CIFAR-10, achieving 89.9% accuracy, significantly surpassing comparably deep conventional CNNs (84.1%) while exhibiting faster convergence and greater training robustness. The core innovation lies in reformulating the learning objective from approximating a target mapping to learning a residual function, synergistically combined with batch normalization. This fundamentally alleviates optimization challenges inherent in very deep networks and establishes a foundational paradigm for deep model architecture design.

0 citationsRead paper

Prompt-and-Check: Using Large Language Models to Evaluate Communication Protocol Compliance in Simulation-Based Training

Aug 12, 2025

In safety-critical domains such as maritime operations, manual assessment of procedural communication compliance suffers from low efficiency and poor reproducibility. To address this, we propose Prompt-and-Check: a zero-shot, context-augmented prompting framework that leverages open-source large language models (LLaMA 2/3, Mistral) on local GPU hardware (RTX 4070) to perform fine-grained compliance classification directly from dialogue transcripts—without model fine-tuning. The method enables context-aware reasoning and fully offline deployment. Experimental evaluation demonstrates strong agreement between model predictions and domain expert annotations (Cohen’s κ > 0.85), significantly enhancing automation and objectivity in post-training debriefing of simulation-based training. Prompt-and-Check establishes a lightweight, interpretable, and deployable paradigm for compliance assessment in high-reliability human–AI collaborative settings.

0 citationsRead paper

Enhancing Egocentric Object Detection in Static Environments using Graph-based Spatial Anomaly Detection and Correction

Aug 11, 2025

To address inconsistent predictions, missed detections, and misclassifications in static first-person object detection—caused by neglecting spatial layout priors—this paper proposes a graph neural network (GNN)-based post-processing framework. It introduces graph-structured modeling for egocentric spatial context for the first time: objects serve as nodes, while geometric and semantic proximity relations define edges, forming a scene graph; GNNs then aggregate neighborhood information to automatically identify and rectify detection anomalies. The method is detector-agnostic, compatible with mainstream detectors such as YOLOv7 and RT-DETR, and requires only minimal human annotation for training. On standard benchmarks, it achieves up to a 4.0% improvement in mAP@50, significantly enhancing detection consistency and robustness under occlusion and cluttered scenes. Key contributions include: (1) the first spatial relational graph modeling paradigm tailored for egocentric vision, and (2) a plug-and-play detection refinement approach that requires no detector retraining.

0 citationsRead paper

Recent publications

Latest Papers

Computer Vision-Based Early Detection of Container Loss at Sea

Apr 27, 2026

This study addresses the critical risk of container stack collapse and loss overboard during vessel motion under harsh sea conditions, which poses significant safety, environmental, and economic threats. To mitigate this challenge, the authors propose a low-cost computer vision system leveraging existing onboard cameras—requiring no additional sensors—that enables container-level micro-motion detection and quantification of relative displacement for the first time. By integrating object segmentation, optical flow-based temporal tracking, and residual motion analysis, the method effectively isolates and monitors inter-tier movements within container stacks in real-world shipborne video footage. The approach significantly enhances cargo safety, operational resilience, and compliance with International Maritime Organization (IMO) regulations, providing a foundation for early-warning systems and timely intervention.

0 citationsRead paper

Eyes on Target: Gaze-Aware Object Detection in Egocentric Video

Nov 03, 2025

To address the neglect of human visual attention in first-person video object detection, this paper proposes a gaze-guided Vision Transformer framework. It models eye-tracking trajectories as dynamic attention priors and injects them into the multi-head self-attention mechanism; a gaze-aware attention head importance metric is further introduced to explicitly modulate each head’s response strength to fixated regions. Additionally, depth information is fused to enhance spatial perception. The method achieves significant improvements over non-gaze-aware baselines—up to +3.2–5.7% mAP—on a custom simulator dataset and public benchmarks including Ego4D Ego-Motion and Ego-CH-Gaze. Ablation studies validate the effectiveness of both the gaze-guidance mechanism and the depth-fusion module. This work is the first to systematically demonstrate the interpretability value of eye movement signals for evaluating first-person object detection, establishing a novel paradigm for attention modeling in embodied intelligence.

0 citationsRead paper

ResNet: Enabling Deep Convolutional Neural Networks through Residual Learning

Oct 27, 2025

To address the training difficulty of ultra-deep convolutional neural networks (CNNs) caused by vanishing gradients, this paper proposes a residual learning framework. It introduces identity-mapping shortcut connections that allow input signals to bypass certain nonlinear transformations, thereby ensuring unimpeded forward signal propagation and backward gradient flow. This approach enables stable end-to-end training of CNNs with over 100 layers—e.g., ResNet-110—on CIFAR-10, achieving 89.9% accuracy, significantly surpassing comparably deep conventional CNNs (84.1%) while exhibiting faster convergence and greater training robustness. The core innovation lies in reformulating the learning objective from approximating a target mapping to learning a residual function, synergistically combined with batch normalization. This fundamentally alleviates optimization challenges inherent in very deep networks and establishes a foundational paradigm for deep model architecture design.

0 citationsRead paper

Prompt-and-Check: Using Large Language Models to Evaluate Communication Protocol Compliance in Simulation-Based Training

Aug 12, 2025

In safety-critical domains such as maritime operations, manual assessment of procedural communication compliance suffers from low efficiency and poor reproducibility. To address this, we propose Prompt-and-Check: a zero-shot, context-augmented prompting framework that leverages open-source large language models (LLaMA 2/3, Mistral) on local GPU hardware (RTX 4070) to perform fine-grained compliance classification directly from dialogue transcripts—without model fine-tuning. The method enables context-aware reasoning and fully offline deployment. Experimental evaluation demonstrates strong agreement between model predictions and domain expert annotations (Cohen’s κ > 0.85), significantly enhancing automation and objectivity in post-training debriefing of simulation-based training. Prompt-and-Check establishes a lightweight, interpretable, and deployable paradigm for compliance assessment in high-reliability human–AI collaborative settings.

0 citationsRead paper

Enhancing Egocentric Object Detection in Static Environments using Graph-based Spatial Anomaly Detection and Correction

Aug 11, 2025

To address inconsistent predictions, missed detections, and misclassifications in static first-person object detection—caused by neglecting spatial layout priors—this paper proposes a graph neural network (GNN)-based post-processing framework. It introduces graph-structured modeling for egocentric spatial context for the first time: objects serve as nodes, while geometric and semantic proximity relations define edges, forming a scene graph; GNNs then aggregate neighborhood information to automatically identify and rectify detection anomalies. The method is detector-agnostic, compatible with mainstream detectors such as YOLOv7 and RT-DETR, and requires only minimal human annotation for training. On standard benchmarks, it achieves up to a 4.0% improvement in mAP@50, significantly enhancing detection consistency and robustness under occlusion and cluttered scenes. Key contributions include: (1) the first spatial relational graph modeling paradigm tailored for egocentric vision, and (2) a plug-and-play detection refinement approach that requires no detector retraining.

0 citationsRead paper