Selected work

Representative Papers

ScratchSim: A Procedural Synthetic Data Pipeline for Surface Scratch Detection

Jul 29, 2026

This study addresses the challenge of surface scratch detection in industrial quality inspection, where performance is often hindered by the scarcity of real annotated data. To overcome this limitation, the authors propose a procedural synthetic data generation pipeline based on BlenderProc, which leverages material modeling, multi-camera configurations, and domain randomization to efficiently produce large-scale scratch images with COCO-format annotations. This approach enables training lightweight detection models—such as YOLOX, YOLOv8, and LW-DETR—without requiring extensive real-world labeled data, thereby facilitating deployment on edge devices. Experimental results demonstrate that fine-tuning models pretrained on synthetic data significantly outperforms training solely on real data across diverse materials. Moreover, when real annotations are extremely limited, a hybrid training strategy effectively recovers detection performance for both convolutional and Transformer-based architectures.

0 citationsRead paper

Human-in-the-Loop Atlas-Based 3D Asset Segmentation for Interactive Content Workflows

Jun 16, 2026

This work addresses the challenge of generating semantic regions for 3D asset segmentation, which traditionally relies on manual intervention and struggles to integrate into interactive content creation pipelines. The authors propose a human-in-the-loop approach for producing editable semantic texture atlases by leveraging multi-view rendering and interactive 2D segmentation—combining SAM² with Label Studio—and back-projecting the results into UV space. A greedy set cover strategy is employed to select key views, enhancing computational efficiency. This method delivers the first unified, editable semantic atlas tailored for XR and game development workflows, enabling downstream tasks such as material assignment and style transfer. Experiments on eight cultural heritage objects demonstrate its effectiveness in handling complex geometries and accurately identifying fine details, cavities, and weak boundaries that require human refinement.

0 citationsRead paper

SynSur: An end-to-end generative pipeline for synthetic industrial surface defect generation and detection

Apr 29, 2026

Industrial defect detection faces significant challenges due to the scarcity, high annotation cost, and imbalanced distribution of real-world labeled data. This work proposes the first end-to-end generative framework that innovatively integrates vision-language model (VLM)-guided prompting, LoRA-finetuned Stable Diffusion, mask-guided inpainting, and automatic filtering with DreamSim and CLIPScore for realistic synthetic defect image generation with precise annotations. Experiments on the BSData and MSD datasets demonstrate that augmenting real data with the generated images during training of detectors such as YOLOv6, YOLOX, and LW-DETR maintains or even slightly improves detection performance, validating the method’s effectiveness and practical utility in cross-domain industrial quality inspection.

0 citationsRead paper

YCB-Ev SD: Synthetic event-vision dataset for 6DoF object pose estimation

Nov 14, 2025

The event-based vision community has long lacked a standardized synthetic dataset for 6DoF object pose estimation, hindering the development of deep learning methods. To address this, we introduce the first synthetic event dataset aligned with the BOP benchmark: high-fidelity event streams are generated via physically based rendering (PBR) and linear motion modeling. We propose a novel dual-channel polarity-encoded time surface representation incorporating linear temporal decay. The dataset comprises 50,000 event sequences, each 34 ms long, spanning multiple objects, viewpoints, and challenging illumination conditions. Experiments demonstrate that our representation significantly outperforms single-channel and exponentially decaying alternatives, yielding a 12.7% improvement in pose estimation accuracy. This work establishes the first standardized synthetic resource for event-based 6DoF pose estimation, providing essential infrastructure for algorithm development and fair, reproducible evaluation.

0 citationsRead paper

Trade-offs in Cross-Domain Generalization of Foundation Model Fine-Tuned for Biometric Applications

Sep 18, 2025

This study identifies a pervasive over-specialization problem in foundational vision-language models (e.g., CLIP) when fine-tuned for biometric tasks—including face recognition, presentation attack detection, and image tampering detection—leading to severe degradation in general visual task performance. Through systematic evaluation across 14 general-purpose vision datasets and multiple biometric benchmarks, we identify task complexity, classifier head design, and model capacity as key determinants of catastrophic forgetting; larger models exhibit greater retention of original generalization capability. We propose the “capacity-mitigates-over-specialization” mechanism and validate it using FRoundation-ViT-L: fine-tuning yields a 58.52% improvement in face recognition accuracy on IJB-C, yet ImageNetV2 accuracy drops to 51.63%—a substantial decline from CLIP’s original 69.84%. This trade-off empirically confirms the fundamental tension between task-specific specialization and cross-domain generalization induced by fine-tuning.

0 citationsRead paper

Recent publications

Latest Papers

ScratchSim: A Procedural Synthetic Data Pipeline for Surface Scratch Detection

Jul 29, 2026

This study addresses the challenge of surface scratch detection in industrial quality inspection, where performance is often hindered by the scarcity of real annotated data. To overcome this limitation, the authors propose a procedural synthetic data generation pipeline based on BlenderProc, which leverages material modeling, multi-camera configurations, and domain randomization to efficiently produce large-scale scratch images with COCO-format annotations. This approach enables training lightweight detection models—such as YOLOX, YOLOv8, and LW-DETR—without requiring extensive real-world labeled data, thereby facilitating deployment on edge devices. Experimental results demonstrate that fine-tuning models pretrained on synthetic data significantly outperforms training solely on real data across diverse materials. Moreover, when real annotations are extremely limited, a hybrid training strategy effectively recovers detection performance for both convolutional and Transformer-based architectures.

0 citationsRead paper

Human-in-the-Loop Atlas-Based 3D Asset Segmentation for Interactive Content Workflows

Jun 16, 2026

This work addresses the challenge of generating semantic regions for 3D asset segmentation, which traditionally relies on manual intervention and struggles to integrate into interactive content creation pipelines. The authors propose a human-in-the-loop approach for producing editable semantic texture atlases by leveraging multi-view rendering and interactive 2D segmentation—combining SAM² with Label Studio—and back-projecting the results into UV space. A greedy set cover strategy is employed to select key views, enhancing computational efficiency. This method delivers the first unified, editable semantic atlas tailored for XR and game development workflows, enabling downstream tasks such as material assignment and style transfer. Experiments on eight cultural heritage objects demonstrate its effectiveness in handling complex geometries and accurately identifying fine details, cavities, and weak boundaries that require human refinement.

0 citationsRead paper

SynSur: An end-to-end generative pipeline for synthetic industrial surface defect generation and detection

Apr 29, 2026

Industrial defect detection faces significant challenges due to the scarcity, high annotation cost, and imbalanced distribution of real-world labeled data. This work proposes the first end-to-end generative framework that innovatively integrates vision-language model (VLM)-guided prompting, LoRA-finetuned Stable Diffusion, mask-guided inpainting, and automatic filtering with DreamSim and CLIPScore for realistic synthetic defect image generation with precise annotations. Experiments on the BSData and MSD datasets demonstrate that augmenting real data with the generated images during training of detectors such as YOLOv6, YOLOX, and LW-DETR maintains or even slightly improves detection performance, validating the method’s effectiveness and practical utility in cross-domain industrial quality inspection.

0 citationsRead paper

YCB-Ev SD: Synthetic event-vision dataset for 6DoF object pose estimation

Nov 14, 2025

The event-based vision community has long lacked a standardized synthetic dataset for 6DoF object pose estimation, hindering the development of deep learning methods. To address this, we introduce the first synthetic event dataset aligned with the BOP benchmark: high-fidelity event streams are generated via physically based rendering (PBR) and linear motion modeling. We propose a novel dual-channel polarity-encoded time surface representation incorporating linear temporal decay. The dataset comprises 50,000 event sequences, each 34 ms long, spanning multiple objects, viewpoints, and challenging illumination conditions. Experiments demonstrate that our representation significantly outperforms single-channel and exponentially decaying alternatives, yielding a 12.7% improvement in pose estimation accuracy. This work establishes the first standardized synthetic resource for event-based 6DoF pose estimation, providing essential infrastructure for algorithm development and fair, reproducible evaluation.

0 citationsRead paper

Trade-offs in Cross-Domain Generalization of Foundation Model Fine-Tuned for Biometric Applications

Sep 18, 2025

This study identifies a pervasive over-specialization problem in foundational vision-language models (e.g., CLIP) when fine-tuned for biometric tasks—including face recognition, presentation attack detection, and image tampering detection—leading to severe degradation in general visual task performance. Through systematic evaluation across 14 general-purpose vision datasets and multiple biometric benchmarks, we identify task complexity, classifier head design, and model capacity as key determinants of catastrophic forgetting; larger models exhibit greater retention of original generalization capability. We propose the “capacity-mitigates-over-specialization” mechanism and validate it using FRoundation-ViT-L: fine-tuning yields a 58.52% improvement in face recognition accuracy on IJB-C, yet ImageNetV2 accuracy drops to 51.63%—a substantial decline from CLIP’s original 69.84%. This trade-off empirically confirms the fundamental tension between task-specific specialization and cross-domain generalization induced by fine-tuning.

0 citationsRead paper