Selected work

Representative Papers

Continuous Speculative Decoding for Autoregressive Image Generation

Nov 18, 2024arXiv.org

Continuous autoregressive visual generation models suffer from high inference latency, while existing speculative decoding methods are restricted to discrete token spaces and lack theoretical foundations or practical techniques for continuous-valued outputs. Method: This work pioneers the extension of speculative decoding to continuous visual generation. We propose a diffusion-prior-based continuous acceptance criterion, design a denoising trajectory alignment mechanism and token pre-filling strategy to mitigate distribution mismatch, and establish a continuous accept-reject sampling framework with analytically derived upper bounds on approximation error. Contribution/Results: Our approach achieves a 2.33× inference speedup on standard diffusion-based autoregressive models while provably preserving the exact output distribution of the original model. The implementation is publicly available.

8 citationsRead paper

Visual Merit or Linguistic Crutch? A Close Look at DeepSeek-OCR

Jan 07, 2026arXiv.org

This study investigates whether DeepSeek-OCR relies on genuine visual capabilities or language priors under high-ratio visual-text compression and evaluates its reliability in long-context scenarios. By employing sentence- and word-level semantic perturbations to disentangle linguistic priors, combined with semantic corruption tests, context stress evaluations, and vision-language decoupling analyses, the work provides the first empirical evidence of the severe dependence of end-to-end OCR models on language priors: accuracy drops precipitously from 90% to 20% without linguistic support, hallucinations increase as visual tokens decrease, and the model completely fails at around 10,000 text tokens. In contrast, traditional OCR methods demonstrate greater robustness. The study also establishes a comprehensive multi-model robustness benchmark encompassing 13 baseline approaches.

1 citations1 influentialRead paper

Micro-expression Recognition Based on Dual-branch Feature Extraction and Fusion

Feb 27, 2026

This work addresses the challenge of micro-expression recognition, which is hindered by their extremely short duration and low intensity, making it difficult for conventional optical flow methods to capture discriminative features effectively. To overcome this limitation, the authors propose a dual-branch feature extraction network that integrates residual structures to mitigate gradient vanishing and Inception modules to enhance multi-scale representation capabilities. Furthermore, a parallel attention mechanism combined with an adaptive feature fusion module is introduced to efficiently integrate multi-source features. Evaluated on the CASME II dataset, the proposed method achieves a recognition accuracy of 74.67%, significantly outperforming LBP-TOP by 11.26% and MSMMT by 3.36%, thereby demonstrating improved robustness and accuracy in micro-expression recognition.

0 citationsRead paper

EPAN: Robust Pedestrian Re-Identification via Enhanced Alignment Network for IoT Surveillance

Nov 03, 2025

To address performance degradation in pedestrian re-identification (ReID) for IoT-based intelligent surveillance—caused by viewpoint variations and environmental interference—this paper proposes the Enhanced Pedestrian Alignment Network (EPAN). EPAN employs a dual-branch architecture integrating multi-scale feature extraction, viewpoint-invariant representation learning, and fine-grained feature alignment to achieve robust cross-camera person representation. Its end-to-end training strategy explicitly mitigates the adverse effects of viewpoint discrepancy and background clutter, significantly enhancing both feature consistency and discriminability. Evaluated on the real-world Inspection-Personnel dataset—specifically constructed for IoT surveillance scenarios—EPAN achieves 90.09% Rank-1 accuracy and 78.82% mean Average Precision (mAP), demonstrating superior effectiveness and practicality in complex, resource-constrained IoT monitoring environments.

0 citationsRead paper

Vision Transformer for Robust Occluded Person Reidentification in Complex Surveillance Scenes

Oct 31, 2025

To address performance degradation in person re-identification (ReID) under complex surveillance scenarios—caused by occlusion, viewpoint distortion, and low-quality imagery—this paper proposes Sh-ViT, a lightweight Vision Transformer (ViT)-based model. Methodologically, it introduces (1) a spatial shuffling module to disrupt ViT’s inherent local inductive bias and enhance robustness against partial occlusion; (2) an adaptive multi-modal data augmentation strategy tailored for surveillance, incorporating geometric transformations, random erasing, blurring, and color jittering; and (3) DeiT-based knowledge distillation for efficient model compression. To alleviate the scarcity of real-world occluded training data, we construct MyTT—the first fine-grained occlusion-specific ReID benchmark. Experiments demonstrate that Sh-ViT achieves 83.2% Rank-1 accuracy and 80.1% mAP on MyTT, and 94.6% Rank-1 accuracy and 87.5% mAP on Market1501—substantially outperforming state-of-the-art CNN- and ViT-based methods.

0 citationsRead paper

Recent publications

Latest Papers

Micro-expression Recognition Based on Dual-branch Feature Extraction and Fusion

Feb 27, 2026

This work addresses the challenge of micro-expression recognition, which is hindered by their extremely short duration and low intensity, making it difficult for conventional optical flow methods to capture discriminative features effectively. To overcome this limitation, the authors propose a dual-branch feature extraction network that integrates residual structures to mitigate gradient vanishing and Inception modules to enhance multi-scale representation capabilities. Furthermore, a parallel attention mechanism combined with an adaptive feature fusion module is introduced to efficiently integrate multi-source features. Evaluated on the CASME II dataset, the proposed method achieves a recognition accuracy of 74.67%, significantly outperforming LBP-TOP by 11.26% and MSMMT by 3.36%, thereby demonstrating improved robustness and accuracy in micro-expression recognition.

0 citationsRead paper

Visual Merit or Linguistic Crutch? A Close Look at DeepSeek-OCR

Jan 07, 2026arXiv.org

This study investigates whether DeepSeek-OCR relies on genuine visual capabilities or language priors under high-ratio visual-text compression and evaluates its reliability in long-context scenarios. By employing sentence- and word-level semantic perturbations to disentangle linguistic priors, combined with semantic corruption tests, context stress evaluations, and vision-language decoupling analyses, the work provides the first empirical evidence of the severe dependence of end-to-end OCR models on language priors: accuracy drops precipitously from 90% to 20% without linguistic support, hallucinations increase as visual tokens decrease, and the model completely fails at around 10,000 text tokens. In contrast, traditional OCR methods demonstrate greater robustness. The study also establishes a comprehensive multi-model robustness benchmark encompassing 13 baseline approaches.

1 citations1 influentialRead paper

EPAN: Robust Pedestrian Re-Identification via Enhanced Alignment Network for IoT Surveillance

Nov 03, 2025

To address performance degradation in pedestrian re-identification (ReID) for IoT-based intelligent surveillance—caused by viewpoint variations and environmental interference—this paper proposes the Enhanced Pedestrian Alignment Network (EPAN). EPAN employs a dual-branch architecture integrating multi-scale feature extraction, viewpoint-invariant representation learning, and fine-grained feature alignment to achieve robust cross-camera person representation. Its end-to-end training strategy explicitly mitigates the adverse effects of viewpoint discrepancy and background clutter, significantly enhancing both feature consistency and discriminability. Evaluated on the real-world Inspection-Personnel dataset—specifically constructed for IoT surveillance scenarios—EPAN achieves 90.09% Rank-1 accuracy and 78.82% mean Average Precision (mAP), demonstrating superior effectiveness and practicality in complex, resource-constrained IoT monitoring environments.

0 citationsRead paper

Vision Transformer for Robust Occluded Person Reidentification in Complex Surveillance Scenes

Oct 31, 2025

To address performance degradation in person re-identification (ReID) under complex surveillance scenarios—caused by occlusion, viewpoint distortion, and low-quality imagery—this paper proposes Sh-ViT, a lightweight Vision Transformer (ViT)-based model. Methodologically, it introduces (1) a spatial shuffling module to disrupt ViT’s inherent local inductive bias and enhance robustness against partial occlusion; (2) an adaptive multi-modal data augmentation strategy tailored for surveillance, incorporating geometric transformations, random erasing, blurring, and color jittering; and (3) DeiT-based knowledge distillation for efficient model compression. To alleviate the scarcity of real-world occluded training data, we construct MyTT—the first fine-grained occlusion-specific ReID benchmark. Experiments demonstrate that Sh-ViT achieves 83.2% Rank-1 accuracy and 80.1% mAP on MyTT, and 94.6% Rank-1 accuracy and 87.5% mAP on Market1501—substantially outperforming state-of-the-art CNN- and ViT-based methods.

0 citationsRead paper

Generating Attribute-Aware Human Motions from Textual Prompt

Jun 27, 2025

Existing text-to-motion generation methods neglect the causal influence of human attributes—such as age, gender, height, and weight—on motion patterns, resulting in generated motions that lack attribute consistency and naturalness. To address this, we propose the first semantic-attribute disentanglement framework grounded in structural causal modeling (SCM), enabling fine-grained, attribute-controllable motion synthesis. We introduce HumanAttr, the first text-motion benchmark dataset annotated with multidimensional, fine-grained human attributes. Our approach integrates semantic-attribute disentangled representation learning, conditional diffusion modeling, and multimodal alignment training. On HumanAttr, our method achieves significant improvements in attribute fidelity (+23.6%) and text-motion alignment (+18.4%). It supports precise, editable attribute control and, for the first time, enables causal disentanglement between human attributes and motion semantics.

0 citationsRead paper