Linlin Yang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Linlin Yang has a total of 67 co-authors. Here are the top co-authors by number of shared publications:
- Angela Yao
National University of Singapore, Singapore, Singapore
7 shared publications Baochang Zhang
5 shared publications- Baochang Zhang
Beihang University
4 shared publications - Kerui Gu
School of Computing, National University of Singapore, Singapore
3 shared publications Boshu Jia
2 shared publicationsBoyu Liu
2 shared publications
Latest Publications
- Open Access
- 31 July 2026
- Published by Elsevier in SSRN Electronic Journal
Bridging the Single-to-Multi-Class Gap in Anomaly Feature Synthesisvia Class-Balanced Patch Sampling
Anomaly-feature synthesis (AFS) has become a simple yet powerful paradigm for anomaly detection: pseudo-anomalies are generated in feature space and a lightweight discriminator is trained to separate them from normal samples. Although AFS methods reach almost perfect AUROC on single-class benchmarks, we show that their performance collapses in multi-class settings. In this paper, we focus on this problem and reveal the primary issue: class-imbalanced feature sampling. When images from many classes are mixed in a mini-batch, dominant classes are oversampled while minority classes are under-represented; pseudo-anomalies are therefore synthesized only around the majority clusters, leaving vast regions of the normal manifold unexplored and rendering the discriminator ineffective. To address this issue, we propose a CoreSet Feature Bank (CFB) framework that:(i) extracts representative patch features per class through patch-wise feature extraction,(ii) builds a cross-class coreset memory that guarantees uniform coverage of the normal manifold,and (iii) synthesizes informative pseudo-anomalies directly within this balanced feature space.By simply replacing the image-level data loader with our patch-level CFB sampler, existing AFS methods can be seamlessly extended to multi-class settings.
- 31 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessBridging the Single-to-Multi-Class Gap in Anomaly Feature Synthesisvia Class-Balanced Patch Sampling
Anomaly-feature synthesis (AFS) has become a simple yet powerful paradigm for anomaly detection: pseudo-anomalies are generated in feature space and a lightweight discriminator is trained to separate them from normal samples. Although AFS methods reach almost perfect AUROC on single-class benchmarks, we show that their performance collapses in multi-class settings. In this paper, we focus on this problem and reveal the primary issue: class-imbalanced feature sampling. When images from many classes are mixed in a mini-batch, dominant classes are oversampled while minority classes are under-represented; pseudo-anomalies are therefore synthesized only around the majority clusters, leaving vast regions of the normal manifold unexplored and rendering the discriminator ineffective. To address this issue, we propose a CoreSet Feature Bank (CFB) framework that:(i) extracts representative patch features per class through patch-wise feature extraction,(ii) builds a cross-class coreset memory that guarantees uniform coverage of the normal manifold,and (iii) synthesizes informative pseudo-anomalies directly within this balanced feature space.By simply replacing the image-level data loader with our patch-level CFB sampler, existing AFS methods can be seamlessly extended to multi-class settings.
- Open Access
- 16 July 2026
- Published by Springer Nature in Research Square
Circuit-Inspired High-Order Neural Networks with Unified Neural Dynamics Modeling for PDE Solving and Visual Perception
Deep networks often rely on architectural heuristics to shape representation evolution, limiting their ability to model data governed by intrinsic dynamics. We present the Circuit-inspired High-Order Neural Network (CHONN), a modular framework that treats representation evolution as a latent potential process and increases its effective order through Kirchhoff-inspired cascade composition. A single Kirchhoff Neural Cell implements a stable first-order update, while serially composed cells form higher-order dynamical operators within one block. This construction is interpretable, numerically stable and compatible with common neural backbones. Theoretical analysis shows that cascaded cells induce end-to-end high-order operators, and controlled experiments demonstrate that intra-block high-order construction differs from generic depth stacking, especially on derivative-sensitive measures. Across steady-state operator learning, long-horizon physical forecasting and ImageNet-1K recognition, CHONN improves structural fidelity, rollout stability and visual representation learning. These results identify high-order circuit composition as a general principle for neural dynamics modeling.
- 16 July 2026
- Published by Springer Nature in Research Square
Open AccessCircuit-Inspired High-Order Neural Networks with Unified Neural Dynamics Modeling for PDE Solving and Visual Perception
Deep networks often rely on architectural heuristics to shape representation evolution, limiting their ability to model data governed by intrinsic dynamics. We present the Circuit-inspired High-Order Neural Network (CHONN), a modular framework that treats representation evolution as a latent potential process and increases its effective order through Kirchhoff-inspired cascade composition. A single Kirchhoff Neural Cell implements a stable first-order update, while serially composed cells form higher-order dynamical operators within one block. This construction is interpretable, numerically stable and compatible with common neural backbones. Theoretical analysis shows that cascaded cells induce end-to-end high-order operators, and controlled experiments demonstrate that intra-block high-order construction differs from generic depth stacking, especially on derivative-sensitive measures. Across steady-state operator learning, long-horizon physical forecasting and ImageNet-1K recognition, CHONN improves structural fidelity, rollout stability and visual representation learning. These results identify high-order circuit composition as a general principle for neural dynamics modeling.
- 30 June 2026
- Published by Elsevier in Neurocomputing
- Vol. 699
HFAT-HMR: Empowering ViT for human mesh recovery via high-frequency enhancement and auxiliary tokens
- 30 June 2026
- Published by Elsevier in Neurocomputing
- Vol. 699
HFAT-HMR: Empowering ViT for human mesh recovery via high-frequency enhancement and auxiliary tokens
- 22 June 2026
- Published by Elsevier in Pattern Recognition
- Vol. 180
Security-aware post-training quantization for Mixture-of-Experts large language models
- 22 June 2026
- Published by Elsevier in Pattern Recognition
- Vol. 180
Security-aware post-training quantization for Mixture-of-Experts large language models
- 2 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. PP (10709908) , 1-5
Industrial Scene Gas Leakage Detection: A Cross-Attention Based Multimodal Feature Difference Network and A New Benchmark
Industrial gas leakage detection is critically important for safety and environmental protection. While infrared imaging enables detection of invisible gases, two challenges remain: existing datasets lack realistic industrial scenarios, and current methods struggle to distinguish gas plumes from background interferences or segment discontinuous gas distributions. This paper introduces a benchmark comprising an Industrial RGB-Thermal Dataset (IRTD) with gas emission and leakage data from laboratory and industrial sites. A VLM-assisted RGBThermal detection framework with a Cross-Attention based Feature Difference (CAFD) module is designed to enhance gasspecific feature differentiation by computing inter-modal feature discrepancies. Evaluations on public datasets and IRTD demonstrate state-of-the-art results.
- 2 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. PP (10709908) , 1-5
Industrial Scene Gas Leakage Detection: A Cross-Attention Based Multimodal Feature Difference Network and A New Benchmark
Industrial gas leakage detection is critically important for safety and environmental protection. While infrared imaging enables detection of invisible gases, two challenges remain: existing datasets lack realistic industrial scenarios, and current methods struggle to distinguish gas plumes from background interferences or segment discontinuous gas distributions. This paper introduces a benchmark comprising an Industrial RGB-Thermal Dataset (IRTD) with gas emission and leakage data from laboratory and industrial sites. A VLM-assisted RGBThermal detection framework with a Cross-Attention based Feature Difference (CAFD) module is designed to enhance gasspecific feature differentiation by computing inter-modal feature discrepancies. Evaluations on public datasets and IRTD demonstrate state-of-the-art results.
- Open Access
- 28 May 2026
- Published by MDPI AG in Journal of Imaging
- Vol. 12 (6) , 237
2s-DAS: Two-Stream Diffusion with Multi-Modal Fusion for Temporal Action Segmentation
Human temporal action segmentation (TAS) is a fundamental video understanding task aimed at partitioning untrimmed videos into semantically coherent action segments. While temporal convolutional networks and transformers have significantly improved frame representation and temporal modeling, existing methods are still constrained by two critical limitations: the dependence on single-modal inputs and the inefficiency of iterative, frame-wise sequential modeling. To address these gaps, we propose 2s-DAS: a novel two-stream diffusion-based framework for action segmentation characterized by three key contributions. First, we introduce a multi-modal frame representation that integrates optical flow with Br-Prompt RGB features, thereby capturing richer spatial-temporal context and enhancing feature representation. Second, we leverage a diffusion model to perform sequence segmentation, utilizing importance sampling to prioritize key frames for segment-level temporal modeling. Concurrently, a refinement mechanism based on iterative decoding denoising is introduced to ensure fine-grained action prediction. Third, we design a two-stream fusion mechanism that processes the streams of RGB with text and optical flow separately and integrates multi-modal information by a late fusion strategy to explicitly reduce oversegmentation. Evaluation experiments on GTEA, 50Salads, and Breakfast datasets show that our 2s-DAS significantly outperforms state-of-the-art methods, setting new benchmarks while effectively addressing the over-segmentation issue.
- 28 May 2026
- Published by MDPI AG in Journal of Imaging
- Vol. 12 (6) , 237
Open Access2s-DAS: Two-Stream Diffusion with Multi-Modal Fusion for Temporal Action Segmentation
Human temporal action segmentation (TAS) is a fundamental video understanding task aimed at partitioning untrimmed videos into semantically coherent action segments. While temporal convolutional networks and transformers have significantly improved frame representation and temporal modeling, existing methods are still constrained by two critical limitations: the dependence on single-modal inputs and the inefficiency of iterative, frame-wise sequential modeling. To address these gaps, we propose 2s-DAS: a novel two-stream diffusion-based framework for action segmentation characterized by three key contributions. First, we introduce a multi-modal frame representation that integrates optical flow with Br-Prompt RGB features, thereby capturing richer spatial-temporal context and enhancing feature representation. Second, we leverage a diffusion model to perform sequence segmentation, utilizing importance sampling to prioritize key frames for segment-level temporal modeling. Concurrently, a refinement mechanism based on iterative decoding denoising is introduced to ensure fine-grained action prediction. Third, we design a two-stream fusion mechanism that processes the streams of RGB with text and optical flow separately and integrates multi-modal information by a late fusion strategy to explicitly reduce oversegmentation. Evaluation experiments on GTEA, 50Salads, and Breakfast datasets show that our 2s-DAS significantly outperforms state-of-the-art methods, setting new benchmarks while effectively addressing the over-segmentation issue.
- 15 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. PP (10518215) , 1
Online Test-time Adaptation for 3D Human Pose Estimation: A Practical Perspective with Noisy 2D Labels
Online Test-Time Adaptation (OTTA) for 3D human pose estimation aims to adapt models to out-of-distribution videos during inference. We focus on sequential adaptation, a setting where frames arrive individually, requiring immediate adaptation on the current frame to obtain reliable estimations before propagating the updated model to the next. While this setting has been proven to achieve optimal performance under the alignment of 2D projection with ground-truths, it often fails in practical scenarios when only noisy 2D labels are available. Specifically, erroneous updates derived from noisy labels inevitably accumulate across videos, and adapting to one frame leads to model overfitting, particularly on challenging poses. This motivates the key challenge of how to exploit informative yet noisy 2D labels while constraining harmful updates during adaptation. To this end, we propose a unified OTTA framework that stabilizes adaptation at both the video and frame levels. At the video level, rather than directly using the propagated model, a robust initialization is reconstructed for each new video. We reset and initialize the model via a confidence-guided experience replay, suppressing accumulated errors while retaining useful prior knowledge. At the frame level, we design a two-stage optimization to separate model propagation from sufficient 2D alignment. An early-stopping signal indicating feature instability is used to maintain a robust model state for propagation to the next frame, while the 2D alignment continues until convergence to provide predictions. This design enhances the adaptation only on the current frame without sacrificing generalization to future frames. Furthermore, we leverage a temporal prior by augmenting nearby confident frames to regularize adaptation on challenging frames. Experiments demonstrate up to 18.0% improvement in Mean Per Joint Position Error over state-of-the-art methods under noisy 2D labels, advancing OTTA towards practical scenarios.
- 15 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. PP (10518215) , 1
Online Test-time Adaptation for 3D Human Pose Estimation: A Practical Perspective with Noisy 2D Labels
Online Test-Time Adaptation (OTTA) for 3D human pose estimation aims to adapt models to out-of-distribution videos during inference. We focus on sequential adaptation, a setting where frames arrive individually, requiring immediate adaptation on the current frame to obtain reliable estimations before propagating the updated model to the next. While this setting has been proven to achieve optimal performance under the alignment of 2D projection with ground-truths, it often fails in practical scenarios when only noisy 2D labels are available. Specifically, erroneous updates derived from noisy labels inevitably accumulate across videos, and adapting to one frame leads to model overfitting, particularly on challenging poses. This motivates the key challenge of how to exploit informative yet noisy 2D labels while constraining harmful updates during adaptation. To this end, we propose a unified OTTA framework that stabilizes adaptation at both the video and frame levels. At the video level, rather than directly using the propagated model, a robust initialization is reconstructed for each new video. We reset and initialize the model via a confidence-guided experience replay, suppressing accumulated errors while retaining useful prior knowledge. At the frame level, we design a two-stage optimization to separate model propagation from sufficient 2D alignment. An early-stopping signal indicating feature instability is used to maintain a robust model state for propagation to the next frame, while the 2D alignment continues until convergence to provide predictions. This design enhances the adaptation only on the current frame without sacrificing generalization to future frames. Furthermore, we leverage a temporal prior by augmenting nearby confident frames to regularize adaptation on challenging frames. Experiments demonstrate up to 18.0% improvement in Mean Per Joint Position Error over state-of-the-art methods under noisy 2D labels, advancing OTTA towards practical scenarios.
- 5 May 2026
- Published by Elsevier in Pattern Recognition
- Vol. 179
LDFE: Laplacian Decoupled Feature Enhancement block for dual-stream CNN-based RGB-IR object detection
- 5 May 2026
- Published by Elsevier in Pattern Recognition
- Vol. 179
LDFE: Laplacian Decoupled Feature Enhancement block for dual-stream CNN-based RGB-IR object detection
- 3 April 2026
- Published by Elsevier in Neurocomputing
- Vol. 684
BinParam: Binarized human parametric modeling via distribution alignment and orthogonal residuals
- 3 April 2026
- Published by Elsevier in Neurocomputing
- Vol. 684
BinParam: Binarized human parametric modeling via distribution alignment and orthogonal residuals
- Open Access
- 27 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. PP (15209210) , 1-12
InstructHumans: Editing Animated 3D Human Textures with Instructions
We present InstructHumans, a novel framework for instruction-driven animatable 3D human texture editing. Existing text-based 3D editing methods often directly apply Score Distillation Sampling (SDS). SDS, designed for generation tasks, cannot account for the defining requirement of editing – maintaining consistency with the source avatar. This work shows that naively using SDS harms editing, as it may destroy consistency. We propose a modified SDS for Editing (SDS-E) that selectively incorporates subterms of SDS across diffusion timesteps. We further enhance SDS-E with spatial smoothness regularization and gradient-based viewpoint sampling for edits with sharp and high-fidelity detailing. Incorporating SDS-E into a 3D human texture editing framework allows us to outperform existing 3D editing methods. Our avatars faithfully reflect the textual edits while remaining consistent with the original avatars. Project page: https://jyzhu.top/instruct-humans/.
- 27 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. PP (15209210) , 1-12
Open AccessInstructHumans: Editing Animated 3D Human Textures with Instructions
We present InstructHumans, a novel framework for instruction-driven animatable 3D human texture editing. Existing text-based 3D editing methods often directly apply Score Distillation Sampling (SDS). SDS, designed for generation tasks, cannot account for the defining requirement of editing – maintaining consistency with the source avatar. This work shows that naively using SDS harms editing, as it may destroy consistency. We propose a modified SDS for Editing (SDS-E) that selectively incorporates subterms of SDS across diffusion timesteps. We further enhance SDS-E with spatial smoothness regularization and gradient-based viewpoint sampling for edits with sharp and high-fidelity detailing. Incorporating SDS-E into a 3D human texture editing framework allows us to outperform existing 3D editing methods. Our avatars faithfully reflect the textual edits while remaining consistent with the original avatars. Project page: https://jyzhu.top/instruct-humans/.