Selected work

Representative Papers

GP-Adapter: Gaussian Process CLIP-Adapter for Few-Shot Out-of-Distribution Detection

Jun 05, 2026

This work addresses the lack of reliable uncertainty estimation in CLIP under few-shot and out-of-distribution (OOD) settings by proposing GP-Adapter, a training-free, plug-and-play framework. GP-Adapter constructs modality-specific one-class Gaussian processes on frozen CLIP features—employing an RBF kernel for images and a linear kernel for text—and fuses predictive statistics from both modalities to produce variance-aware confidence scores. As the first approach to integrate Gaussian process-based uncertainty modeling into CLIP, GP-Adapter enables effective few-shot classification and OOD detection without fine-tuning, incurs low memory overhead, and complements existing prompt-learning methods. Experiments demonstrate competitive classification performance on ImageNet and multiple OOD benchmarks, along with significantly improved OOD detection accuracy.

0 citationsRead paper

JOPP-3D: Joint Open Vocabulary Semantic Segmentation on Point Clouds and Panoramas

Mar 06, 2026

This work addresses the challenges of scarce annotated data and limited generalization of fixed-label models in cross-modal (3D point clouds and panoramic images) semantic segmentation by proposing the first open-vocabulary semantic segmentation framework driven by language for cross-modal scene understanding. The method converts RGB-D panoramic images into tangent-plane views and aligns them with 3D point clouds, enabling joint extraction of features from vision-language foundation models to achieve cross-modal semantic alignment and consistent segmentation. The framework supports generating semantic masks from natural language queries and significantly outperforms existing methods on the Stanford-2D-3D-s and ToF-360 datasets, achieving state-of-the-art performance under both open- and closed-vocabulary settings.

0 citationsRead paper

RegMean++: Enhancing Effectiveness and Generalization of Regression Mean for Model Merging

Aug 05, 2025

RegMean independently optimizes individual linear layers during model merging, neglecting inter-layer feature propagation and dependencies, thereby inadequately modeling merged-model behavior. To address this, we propose RegMean++, the first method within the regression-mean framework to explicitly model intra-layer consistency and cross-layer dependencies. RegMean++ integrates feature propagation analysis with linear regression and employs a closed-form solution for efficient, interpretable weight optimization. Experimental results demonstrate that RegMean++ significantly improves prediction consistency across diverse evaluation settings: it outperforms RegMean in both in-distribution and out-of-distribution generalization, scales effectively to large-scale tasks, and exhibits robustness under distributional shift. Notably, RegMean++ achieves state-of-the-art performance while retaining computational efficiency and analytical transparency.

0 citationsRead paper

Guidelines for External Disturbance Factors in the Use of OCR in Real-World Environments

Apr 21, 2025

External disturbances—such as illumination variation, occlusion, and geometric distortion—degrade OCR input images in real-world scenarios, severely compromising recognition accuracy and complicating quality control. To address this, we propose the first standardized taxonomy of external disturbance factors specifically designed for OCR robustness evaluation, systematically categorizing disturbance types and their associated image degradation patterns. Our methodology integrates empirical analysis, cross-scenario degradation modeling, error attribution, and engineering validation to establish a comprehensive assessment framework. We deliver a structured disturbance factor table and actionable guidelines for OCR deployment and quality control. This work bridges the gap between laboratory-reported OCR performance and real-world reliability, significantly improving deployment success rates and quality controllability of OCR systems under complex environmental conditions.

0 citationsRead paper

Reinforcement Learning of Flexible Policies for Symbolic Instructions With Adjustable Mapping Specifications

Jan 31, 2025IEEE Robotics and Automation Letters

To address the challenge of dynamically interpreting identical symbolic instructions based on contextual state in multi-view robotic equipment inspection, this paper proposes a reinforcement learning framework that decouples symbolic semantics from mapping specifications. Methodologically, it (1) explicitly separates symbolic instruction representation—formalized in Linear Temporal Logic—from learnable mapping specifications; (2) introduces a specification-aware state modulation mechanism to enable state-dependent semantic adaptation; and (3) designs a symbol-cardinality-driven task curriculum learning strategy for progressive complexity scaling. Evaluated in a 3D simulation environment, the approach achieves a 27.4% higher success rate on flexible mapping tasks compared to a context-aware multitask RL baseline, demonstrating significantly improved generalization and environmental robustness.

0 citationsRead paper

Recent publications

Latest Papers

GP-Adapter: Gaussian Process CLIP-Adapter for Few-Shot Out-of-Distribution Detection

Jun 05, 2026

This work addresses the lack of reliable uncertainty estimation in CLIP under few-shot and out-of-distribution (OOD) settings by proposing GP-Adapter, a training-free, plug-and-play framework. GP-Adapter constructs modality-specific one-class Gaussian processes on frozen CLIP features—employing an RBF kernel for images and a linear kernel for text—and fuses predictive statistics from both modalities to produce variance-aware confidence scores. As the first approach to integrate Gaussian process-based uncertainty modeling into CLIP, GP-Adapter enables effective few-shot classification and OOD detection without fine-tuning, incurs low memory overhead, and complements existing prompt-learning methods. Experiments demonstrate competitive classification performance on ImageNet and multiple OOD benchmarks, along with significantly improved OOD detection accuracy.

0 citationsRead paper

JOPP-3D: Joint Open Vocabulary Semantic Segmentation on Point Clouds and Panoramas

Mar 06, 2026

This work addresses the challenges of scarce annotated data and limited generalization of fixed-label models in cross-modal (3D point clouds and panoramic images) semantic segmentation by proposing the first open-vocabulary semantic segmentation framework driven by language for cross-modal scene understanding. The method converts RGB-D panoramic images into tangent-plane views and aligns them with 3D point clouds, enabling joint extraction of features from vision-language foundation models to achieve cross-modal semantic alignment and consistent segmentation. The framework supports generating semantic masks from natural language queries and significantly outperforms existing methods on the Stanford-2D-3D-s and ToF-360 datasets, achieving state-of-the-art performance under both open- and closed-vocabulary settings.

0 citationsRead paper

RegMean++: Enhancing Effectiveness and Generalization of Regression Mean for Model Merging

Aug 05, 2025

RegMean independently optimizes individual linear layers during model merging, neglecting inter-layer feature propagation and dependencies, thereby inadequately modeling merged-model behavior. To address this, we propose RegMean++, the first method within the regression-mean framework to explicitly model intra-layer consistency and cross-layer dependencies. RegMean++ integrates feature propagation analysis with linear regression and employs a closed-form solution for efficient, interpretable weight optimization. Experimental results demonstrate that RegMean++ significantly improves prediction consistency across diverse evaluation settings: it outperforms RegMean in both in-distribution and out-of-distribution generalization, scales effectively to large-scale tasks, and exhibits robustness under distributional shift. Notably, RegMean++ achieves state-of-the-art performance while retaining computational efficiency and analytical transparency.

0 citationsRead paper

Guidelines for External Disturbance Factors in the Use of OCR in Real-World Environments

Apr 21, 2025

External disturbances—such as illumination variation, occlusion, and geometric distortion—degrade OCR input images in real-world scenarios, severely compromising recognition accuracy and complicating quality control. To address this, we propose the first standardized taxonomy of external disturbance factors specifically designed for OCR robustness evaluation, systematically categorizing disturbance types and their associated image degradation patterns. Our methodology integrates empirical analysis, cross-scenario degradation modeling, error attribution, and engineering validation to establish a comprehensive assessment framework. We deliver a structured disturbance factor table and actionable guidelines for OCR deployment and quality control. This work bridges the gap between laboratory-reported OCR performance and real-world reliability, significantly improving deployment success rates and quality controllability of OCR systems under complex environmental conditions.

0 citationsRead paper

Reinforcement Learning of Flexible Policies for Symbolic Instructions With Adjustable Mapping Specifications

Jan 31, 2025IEEE Robotics and Automation Letters

To address the challenge of dynamically interpreting identical symbolic instructions based on contextual state in multi-view robotic equipment inspection, this paper proposes a reinforcement learning framework that decouples symbolic semantics from mapping specifications. Methodologically, it (1) explicitly separates symbolic instruction representation—formalized in Linear Temporal Logic—from learnable mapping specifications; (2) introduces a specification-aware state modulation mechanism to enable state-dependent semantic adaptation; and (3) designs a symbol-cardinality-driven task curriculum learning strategy for progressive complexity scaling. Evaluated in a 3D simulation environment, the approach achieves a 27.4% higher success rate on flexible mapping tasks compared to a context-aware multitask RL baseline, demonstrating significantly improved generalization and environmental robustness.

0 citationsRead paper