Aidong Men
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Aidong Men has a total of 69 co-authors. Here are the top co-authors by number of shared publications:
- Qingchao Chen
Peking University
5 shared publications - Zhuqing Jiang
School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing, China
5 shared publications - Haiyong Luo
Institute of Computing Technology, Chinese Academy of Sciences
4 shared publications - Jiangbo Pei
National Institute of Health Data Science, Peking University, Beijing, China
4 shared publications - Qu Wang
School of Automation Science and Electrical Engineering, University of Science and Technology Beijing, Beijing 100083, China
4 shared publications - Fang Zhao
Beijing University of Posts and Telecommunications
3 shared publications
Latest Publications
- 29 December 2025
- Published by Elsevier in Knowledge-Based Systems
- Vol. 335
Zero-shot diverse audio captioning with diffusion models
- 29 December 2025
- Published by Elsevier in Knowledge-Based Systems
- Vol. 335
Zero-shot diverse audio captioning with diffusion models
- Open Access
Boosting Diffusion Networks with Deep External Context-Aware Encoders for Low-Light Image Enhancement
Low-light image enhancement (LLIE) requires modeling spatially extensive and interdependent degradations across large pixel regions, while directly equipping diffusion-based LLIE with heavy global modules inside the iterative denoising backbone leads to prohibitive computational overhead. To enhance long-range context modeling without inflating the per-step cost of diffusion, we propose ECA-Diff, a diffusion framework augmented with a deep External Context-Aware Encoder (ECAE). A latent-space context network built with hybrid Transformer–Convolution blocks extracts holistic cues from the input, generates multi-scale context features once, and injects them into the diffusion backbone as lightweight conditional guidance across all sampling steps. In addition, a CIELAB-space Luminance-Adaptive Chromaticity Loss regularizes conditional diffusion training and mitigates the cool color cast frequently observed in low-luminance regions. Experiments on paired and unpaired benchmarks show that ECA-Diff consistently outperforms recent state-of-the-art LLIE methods in both full-reference (PSNR/SSIM/LPIPS) and no-reference (NIQE/BRISQUE) metrics, with the external context path introducing only modest overhead relative to the baseline diffusion backbone. These results indicate that decoupling global context estimation from the iterative denoising process is an effective way to boost diffusion-based LLIE and provides a general compute-once conditioning paradigm for low-level image restoration.
Open AccessBoosting Diffusion Networks with Deep External Context-Aware Encoders for Low-Light Image Enhancement
Low-light image enhancement (LLIE) requires modeling spatially extensive and interdependent degradations across large pixel regions, while directly equipping diffusion-based LLIE with heavy global modules inside the iterative denoising backbone leads to prohibitive computational overhead. To enhance long-range context modeling without inflating the per-step cost of diffusion, we propose ECA-Diff, a diffusion framework augmented with a deep External Context-Aware Encoder (ECAE). A latent-space context network built with hybrid Transformer–Convolution blocks extracts holistic cues from the input, generates multi-scale context features once, and injects them into the diffusion backbone as lightweight conditional guidance across all sampling steps. In addition, a CIELAB-space Luminance-Adaptive Chromaticity Loss regularizes conditional diffusion training and mitigates the cool color cast frequently observed in low-luminance regions. Experiments on paired and unpaired benchmarks show that ECA-Diff consistently outperforms recent state-of-the-art LLIE methods in both full-reference (PSNR/SSIM/LPIPS) and no-reference (NIQE/BRISQUE) metrics, with the external context path introducing only modest overhead relative to the baseline diffusion backbone. These results indicate that decoupling global context estimation from the iterative denoising process is an effective way to boost diffusion-based LLIE and provides a general compute-once conditioning paradigm for low-level image restoration.
- 31 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 47  (11) , 10112-10128
Selection, Ensemble, and Adaptation: Advancing Multi-Source-Free Domain Adaptation via Architecture Zoo
Conventional Multi-Source Free Domain Adaptation (MSFDA) assumes that each source domain provides a single source model, and all source models adopt a uniform architecture. This paper introduces Zoo-MSFDA, a more general setting that allows each source domain to offer a zoo of multiple source models with different architectures. While it enriches the source knowledge, Zoo-MSFDA risks being dominated by suboptimal/harmful models. To address this issue, we theoretically analyze the model selection problem in Zoo-MSFDA, and introduce two principles: transferability principle and diversity principle. Recognizing the challenge of measuring transferability, we subsequently propose a novel Source-Free Unsupervised Transferability Estimation (SUTE). It enables assessing and comparing transferability across multiple source models with different architectures under domain shift, without requiring target labels and source data. Based on above, we introduce a Selection, Ensemble, and Adaptation (SEA) framework to address Zoo-MSFDA, which consists of: 1) source models selection based on the proposed principles and SUTE; 2) ensemble construction based on SUTE-estimated transferability; 3) target-domain adaptation of the ensemble model. Evaluations demonstrate that our SEA framework, with the introduced Zoo-MSFDA setting, significantly improves adaptation performance in 2D image classification tasks. Additionally, our SUTE achieves state-of-the-art performance in transferability estimation.
- 31 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 47  (11) , 10112-10128
Selection, Ensemble, and Adaptation: Advancing Multi-Source-Free Domain Adaptation via Architecture Zoo
Conventional Multi-Source Free Domain Adaptation (MSFDA) assumes that each source domain provides a single source model, and all source models adopt a uniform architecture. This paper introduces Zoo-MSFDA, a more general setting that allows each source domain to offer a zoo of multiple source models with different architectures. While it enriches the source knowledge, Zoo-MSFDA risks being dominated by suboptimal/harmful models. To address this issue, we theoretically analyze the model selection problem in Zoo-MSFDA, and introduce two principles: transferability principle and diversity principle. Recognizing the challenge of measuring transferability, we subsequently propose a novel Source-Free Unsupervised Transferability Estimation (SUTE). It enables assessing and comparing transferability across multiple source models with different architectures under domain shift, without requiring target labels and source data. Based on above, we introduce a Selection, Ensemble, and Adaptation (SEA) framework to address Zoo-MSFDA, which consists of: 1) source models selection based on the proposed principles and SUTE; 2) ensemble construction based on SUTE-estimated transferability; 3) target-domain adaptation of the ensemble model. Evaluations demonstrate that our SEA framework, with the introduced Zoo-MSFDA setting, significantly improves adaptation performance in 2D image classification tasks. Additionally, our SUTE achieves state-of-the-art performance in transferability estimation.
- 23 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 47  (9) , 7930-7943
Incorporating Pre-Training Data Matters in Unsupervised Domain Adaptation
In deep learning, initializing models with pre-trained weights has become the de facto practice for various downstream tasks. Many unsupervised domain adaptation (UDA) methods typically adopt a backbone pre-trained on ImageNet, and focus on reducing the source-target domain discrepancy. However, the impact of pre-training on adaptation received little attention. In this study, we delve into UDA from the novel perspective of pre-training. We first demonstrate the impact of pre-training by analyzing the dynamic distribution discrepancies between pre-training data domain and the source/ target domain during adaptation. Then, we reveal that the target error also stems from the pre-training in the following two factors: 1) empirically, target error arises from the gradually degenerative pre-trained knowledge during adaptation; 2) theoretically, the error bound depends on difference between the gradient of loss function, i.e., on the target domain and pre-training data domain. To address these two issues, we redefine UDA as a three-domain problem, i.e., source domain, target domain, and pre-training data domain; then we propose a novel framework, named TriDA. We maintain the pre-trained knowledge and improve the error bound by incorporating pre-training data into adaptation for both vanilla UDA and source-free UDA scenarios. For efficiency, we introduce a selection strategy for pre-training data, and offer a solution with synthesized images when pre-training data is unavailable during adaptation. Notably, TriDA is effective even with a small amount of pre-training or synthesized images, and seamlessly complements the two scenario UDA methods, demonstrating state-of-the-art performance across multiple benchmarks. We hope our work provides new insights for better understanding and application of domain adaptation.
- 23 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 47  (9) , 7930-7943
Incorporating Pre-Training Data Matters in Unsupervised Domain Adaptation
In deep learning, initializing models with pre-trained weights has become the de facto practice for various downstream tasks. Many unsupervised domain adaptation (UDA) methods typically adopt a backbone pre-trained on ImageNet, and focus on reducing the source-target domain discrepancy. However, the impact of pre-training on adaptation received little attention. In this study, we delve into UDA from the novel perspective of pre-training. We first demonstrate the impact of pre-training by analyzing the dynamic distribution discrepancies between pre-training data domain and the source/ target domain during adaptation. Then, we reveal that the target error also stems from the pre-training in the following two factors: 1) empirically, target error arises from the gradually degenerative pre-trained knowledge during adaptation; 2) theoretically, the error bound depends on difference between the gradient of loss function, i.e., on the target domain and pre-training data domain. To address these two issues, we redefine UDA as a three-domain problem, i.e., source domain, target domain, and pre-training data domain; then we propose a novel framework, named TriDA. We maintain the pre-trained knowledge and improve the error bound by incorporating pre-training data into adaptation for both vanilla UDA and source-free UDA scenarios. For efficiency, we introduce a selection strategy for pre-training data, and offer a solution with synthesized images when pre-training data is unavailable during adaptation. Notably, TriDA is effective even with a small amount of pre-training or synthesized images, and seamlessly complements the two scenario UDA methods, demonstrating state-of-the-art performance across multiple benchmarks. We hope our work provides new insights for better understanding and application of domain adaptation.
- 21 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12  (12) , 22381-22392
Camera-Invariant Meta-Learning Network for Single-Camera-Training Person Reidentification
Single-camera-training person re-identification (SCT re-ID) aims to train a re-ID model using SCT datasets where each person appears in only one camera. The main challenge of SCT re-ID is to learn camera-invariant feature representations without cross-camera same-person (CCSP) data as supervision. Previous methods address it by assuming that the most similar person should be found in another camera. However, this assumption is not guaranteed to be correct. In this paper, we propose a novel solution: the Camera-Invariant Meta-Learning Network (CIMN) for SCT re-ID. CIMN operates under the premise that camera-invariant feature representations should remain robust despite changes in camera settings. To achieve this, we partition the training data into a meta-train set and a meta-test set based on camera IDs. We then conduct a cross-camera simulation using a meta-learning strategy, aiming to enforce the feature representations learned from the meta-train set to be robust when applied to the meta-test set. We further introduce three specific loss functions to leverage potential identity relations between the meta-train set and the meta-test set. Through the cross-camera simulation and the introduced loss functions, CIMN can extract feature representations that are both camera-invariant and identity-discriminative even in the absence of CCSP data. Our experimental results demonstrate that CIMN can extract feature representations that are both camera-invariant and identity-discriminative, even in the absence of CCSP data. our method achieves comparable performance with and without the use of CCSP data, and outperforms state-of-the-art methods on three SCT re-ID benchmarks.
- 21 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12  (12) , 22381-22392
Camera-Invariant Meta-Learning Network for Single-Camera-Training Person Reidentification
Single-camera-training person re-identification (SCT re-ID) aims to train a re-ID model using SCT datasets where each person appears in only one camera. The main challenge of SCT re-ID is to learn camera-invariant feature representations without cross-camera same-person (CCSP) data as supervision. Previous methods address it by assuming that the most similar person should be found in another camera. However, this assumption is not guaranteed to be correct. In this paper, we propose a novel solution: the Camera-Invariant Meta-Learning Network (CIMN) for SCT re-ID. CIMN operates under the premise that camera-invariant feature representations should remain robust despite changes in camera settings. To achieve this, we partition the training data into a meta-train set and a meta-test set based on camera IDs. We then conduct a cross-camera simulation using a meta-learning strategy, aiming to enforce the feature representations learned from the meta-train set to be robust when applied to the meta-test set. We further introduce three specific loss functions to leverage potential identity relations between the meta-train set and the meta-test set. Through the cross-camera simulation and the introduced loss functions, CIMN can extract feature representations that are both camera-invariant and identity-discriminative even in the absence of CCSP data. Our experimental results demonstrate that CIMN can extract feature representations that are both camera-invariant and identity-discriminative, even in the absence of CCSP data. our method achieves comparable performance with and without the use of CCSP data, and outperforms state-of-the-art methods on three SCT re-ID benchmarks.
- Open Access
- 5 March 2025
- Published by MDPI AG in Brain Sciences
- Vol. 15  (3) , 277
Evaluating Cognitive Function and Brain Activity Patterns via Blood Oxygen Level-Dependent Transformer in N-Back Working Memory Tasks
(1) Background: Working memory, which involves temporary storage, information processing, and regulating attention resources, is a fundamental cognitive process and constitutes a significant component of neuroscience research. This study aimed to evaluate brain activation patterns by analyzing functional magnetic resonance imaging (fMRI) time-series data collected during a designed N-back working memory task with varying cognitive demands. (2) Methods: We utilized a novel transformer model, blood oxygen level-dependent transformer (BolT), to extract the activation level features of brain regions in the cognitive process, thereby obtaining the influence weights of regions of interest (ROIs) on the corresponding tasks. (3) Results: Compared with previous studies, our work reached similar conclusions in major brain region performance and provides a more precise analysis for identifying brain activation patterns. For each type of working memory task, we selected the top 5 percent of the most influential ROIs and conducted a comprehensive analysis and discussion. Additionally, we explored the effect of prior knowledge conditions on the performance of different tasks in the same period and the same tasks at different times. (4) Conclusions: The comparison results reflect the brain’s adaptive strategies and dependencies in coping with different levels of cognitive demands and the stability optimization of the brain’s cognitive processing. This study introduces innovative methodologies for understanding brain function and cognitive processes, highlighting the potential of transformer in cognitive neuroscience. Its findings offer new insights into brain activity patterns associated with working memory, contributing to the broader landscape of neuroscience research.
- 5 March 2025
- Published by MDPI AG in Brain Sciences
- Vol. 15  (3) , 277
Open AccessEvaluating Cognitive Function and Brain Activity Patterns via Blood Oxygen Level-Dependent Transformer in N-Back Working Memory Tasks
(1) Background: Working memory, which involves temporary storage, information processing, and regulating attention resources, is a fundamental cognitive process and constitutes a significant component of neuroscience research. This study aimed to evaluate brain activation patterns by analyzing functional magnetic resonance imaging (fMRI) time-series data collected during a designed N-back working memory task with varying cognitive demands. (2) Methods: We utilized a novel transformer model, blood oxygen level-dependent transformer (BolT), to extract the activation level features of brain regions in the cognitive process, thereby obtaining the influence weights of regions of interest (ROIs) on the corresponding tasks. (3) Results: Compared with previous studies, our work reached similar conclusions in major brain region performance and provides a more precise analysis for identifying brain activation patterns. For each type of working memory task, we selected the top 5 percent of the most influential ROIs and conducted a comprehensive analysis and discussion. Additionally, we explored the effect of prior knowledge conditions on the performance of different tasks in the same period and the same tasks at different times. (4) Conclusions: The comparison results reflect the brain’s adaptive strategies and dependencies in coping with different levels of cognitive demands and the stability optimization of the brain’s cognitive processing. This study introduces innovative methodologies for understanding brain function and cognitive processes, highlighting the potential of transformer in cognitive neuroscience. Its findings offer new insights into brain activity patterns associated with working memory, contributing to the broader landscape of neuroscience research.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 12384-12395
Socratic Style Chain-of-Thoughts Help LLMs to be a Better Reasoner
Jiangbo Pei, Peiyu Liu, Xin Zhao, Aidong Men, Yang Liu. Findings of the Association for Computational Linguistics: ACL 2025. 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 12384-12395
Open AccessSocratic Style Chain-of-Thoughts Help LLMs to be a Better Reasoner
Jiangbo Pei, Peiyu Liu, Xin Zhao, Aidong Men, Yang Liu. Findings of the Association for Computational Linguistics: ACL 2025. 2025.
- Open Access
- 29 November 2024
- Published by MDPI AG in Big Data and Cognitive Computing
- Vol. 8  (12) , 175
Semi-Open Set Object Detection Algorithm Leveraged by Multi-Modal Large Language Models
Currently, closed-set object detection models represented by YOLO are widely deployed in the industrial field. However, such closed-set models lack sufficient tuning ability for easily confused objects in complex detection scenarios. Open-set object detection models such as GroundingDINO expand the detection range to a certain extent, but they still have a gap in detection accuracy compared with closed-set detection models and cannot meet the requirements for high-precision detection in practical applications. In addition, existing detection technologies are also insufficient in interpretability, making it difficult to clearly show users the basis and process of judgment of detection results, causing users to have doubts about the trust and application of detection results. Based on the above deficiencies, we propose a new object detection algorithm based on multi-modal large language models that significantly improves the detection effect of closed-set object detection models for more difficult boundary tasks while ensuring detection accuracy, thereby achieving a semi-open set object detection algorithm. It has significant improvements in accuracy and interpretability under the verification of seven common traffic and safety production scenarios.
- 29 November 2024
- Published by MDPI AG in Big Data and Cognitive Computing
- Vol. 8  (12) , 175
Open AccessSemi-Open Set Object Detection Algorithm Leveraged by Multi-Modal Large Language Models
Currently, closed-set object detection models represented by YOLO are widely deployed in the industrial field. However, such closed-set models lack sufficient tuning ability for easily confused objects in complex detection scenarios. Open-set object detection models such as GroundingDINO expand the detection range to a certain extent, but they still have a gap in detection accuracy compared with closed-set detection models and cannot meet the requirements for high-precision detection in practical applications. In addition, existing detection technologies are also insufficient in interpretability, making it difficult to clearly show users the basis and process of judgment of detection results, causing users to have doubts about the trust and application of detection results. Based on the above deficiencies, we propose a new object detection algorithm based on multi-modal large language models that significantly improves the detection effect of closed-set object detection models for more difficult boundary tasks while ensuring detection accuracy, thereby achieving a semi-open set object detection algorithm. It has significant improvements in accuracy and interpretability under the verification of seven common traffic and safety production scenarios.
- 22 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 33  (10577149) , 6204-6215
EviPrompt: A Training-Free Evidential Prompt Generation Method for Adapting Segment Anything Model in Medical Images
Medical image segmentation is a critical task in clinical applications. Recently, the Segment Anything Model (SAM) has demonstrated potential for natural image segmentation. However, the requirement for expert labour to provide prompts, and the domain gap between natural and medical images pose significant obstacles in adapting SAM to medical images. To overcome these challenges, this paper introduces a novel prompt generation method named EviPrompt. The proposed method requires only a single reference image-annotation pair, making it a training-free solution that significantly reduces the need for extensive labelling and computational resources. First, prompts are automatically generated based on the similarity between features of the reference and target images, and evidential learning is introduced to improve reliability. Then, to mitigate the impact of the domain gap, committee voting and inference-guided in-context learning are employed, generating prompts primarily based on human prior knowledge and reducing reliance on extracted semantic information. EviPrompt represents an efficient and robust approach to medical image segmentation. We evaluate it across a broad range of tasks and modalities, confirming its efficacy. The source code is available at https://github.com/SPIresearch/EviPrompt.
- 22 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 33  (10577149) , 6204-6215
EviPrompt: A Training-Free Evidential Prompt Generation Method for Adapting Segment Anything Model in Medical Images
Medical image segmentation is a critical task in clinical applications. Recently, the Segment Anything Model (SAM) has demonstrated potential for natural image segmentation. However, the requirement for expert labour to provide prompts, and the domain gap between natural and medical images pose significant obstacles in adapting SAM to medical images. To overcome these challenges, this paper introduces a novel prompt generation method named EviPrompt. The proposed method requires only a single reference image-annotation pair, making it a training-free solution that significantly reduces the need for extensive labelling and computational resources. First, prompts are automatically generated based on the similarity between features of the reference and target images, and evidential learning is introduced to improve reliability. Then, to mitigate the impact of the domain gap, committee voting and inference-guided in-context learning are employed, generating prompts primarily based on human prior knowledge and reducing reliance on extracted semantic information. EviPrompt represents an efficient and robust approach to medical image segmentation. We evaluate it across a broad range of tasks and modalities, confirming its efficacy. The source code is available at https://github.com/SPIresearch/EviPrompt.
- 1 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 11  (15) , 26020-26032
Pedestrian Navigation Activity Recognition Based on Segmentation Transformer
In the context of the Internet of Things, utilizing the inherent inertial sensors in smartphones for human activity recognition (HAR) has garnered considerable attention owing to its wide-ranging applications. However, prevailing HAR approaches primarily treat activity identification as a single-label classification task, focusing solely on discerning pedestrian motion modes or device usage modes, while disregarding their interrelatedness. Additionally, HAR methods employing sliding windows encounter challenges associated with the multiclass window problem, wherein certain sample labels differ from the label assigned to the window. This article aims to address these issues. This article presents a novel approach for simultaneously recognizing pedestrian motion and device usage modes by utilizing the segmentation Transformer. The proposed joint recognition framework effectively annotates sensor data at each timestamp and achieves dense prediction of time-series data through the encoding and decoding of the annotated data. To optimize the utilization of information extracted from each Transformer layer, a global up-sampling decoder based on the pyramid attention module is introduced, enabling dense decoding of features obtained from each Transformer layer. We performed experiments on two publicly available data sets to comprehensively assess the effectiveness of the proposed methodology. The results demonstrate that our approach achieves an accuracy of 99.79% and a weighted F-score of 99.77%, surpassing the performance of existing state-of-the-art methods. Furthermore, we constructed heterogeneous data sets to validate the robustness of our method. The extensive experimental findings indicate that the joint recognition framework effectively uncovers the inherent correlations between pedestrian motion and device usage modes, leading to enhanced accuracy in recognition and addressing the challenges posed by the multiclass window problem.
- 1 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 11  (15) , 26020-26032
Pedestrian Navigation Activity Recognition Based on Segmentation Transformer
In the context of the Internet of Things, utilizing the inherent inertial sensors in smartphones for human activity recognition (HAR) has garnered considerable attention owing to its wide-ranging applications. However, prevailing HAR approaches primarily treat activity identification as a single-label classification task, focusing solely on discerning pedestrian motion modes or device usage modes, while disregarding their interrelatedness. Additionally, HAR methods employing sliding windows encounter challenges associated with the multiclass window problem, wherein certain sample labels differ from the label assigned to the window. This article aims to address these issues. This article presents a novel approach for simultaneously recognizing pedestrian motion and device usage modes by utilizing the segmentation Transformer. The proposed joint recognition framework effectively annotates sensor data at each timestamp and achieves dense prediction of time-series data through the encoding and decoding of the annotated data. To optimize the utilization of information extracted from each Transformer layer, a global up-sampling decoder based on the pyramid attention module is introduced, enabling dense decoding of features obtained from each Transformer layer. We performed experiments on two publicly available data sets to comprehensively assess the effectiveness of the proposed methodology. The results demonstrate that our approach achieves an accuracy of 99.79% and a weighted F-score of 99.77%, surpassing the performance of existing state-of-the-art methods. Furthermore, we constructed heterogeneous data sets to validate the robustness of our method. The extensive experimental findings indicate that the joint recognition framework effectively uncovers the inherent correlations between pedestrian motion and device usage modes, leading to enhanced accuracy in recognition and addressing the challenges posed by the multiclass window problem.