Duc Tai Phan

denotes journal

denotes conference

2026

1. ICIT’26

   

   Three-Rank Calibrated Entropy-Weighted Clustering for Active Learning with Vision-Language Models

   Duc Tai Phan, Lam Phuong Nguyen, Nam Van Hai Phan, Tri Nguyen Thai, and Duc Ngoc Minh Dang

   In The 5th International Conference on Intelligence of Things 2026 (ICITconf’2026) , Seoul, South Korea, Oct 2026

   Abs Bib Code

   Vision-language models (VLMs) provide strong transferable representations for visual recognition, but adapting them to downstream tasks remains challenging when labeled data are scarce. Active learning (AL) addresses this issue by selecting informative unlabeled samples for annotation; however, recent VLM-oriented solutions, such as Calibrated Entropy-weighted Clustering (CEC), select the final query from each cluster using a nearest-center rule. This rule is representative, but it does not consider whether a candidate is predicted to belong to a class with few labeled examples. To address this limitation, we propose Three-Rank Calibrated Entropy-weighted Clustering (R3-CEC), a parameter-free extension of CEC that replaces the final within-cluster selection step with rank aggregation over residual class coverage, informativeness, and representativeness. Experiments on six image classification benchmarks across multiple prompt-learning baselines show that R3-CEC improves over CEC in the main setting across multiple budget settings, with gains of +3.2%, +1.4%, and +0.2%. These results indicate that R3-CEC is a simple and practical query strategy for annotation-efficient VLM adaptation, with potential applications in label-scarce visual recognition scenarios.

   @inproceedings{Phan2610R3CEC,
     author = {Phan, Duc Tai and Nguyen, Lam Phuong and Phan, Nam Van Hai and Thai, Tri Nguyen and Dang, Duc Ngoc Minh},
     title = {{Three-Rank} Calibrated {Entropy-Weighted} Clustering for Active Learning with {Vision-Language} Models},
     booktitle = {The 5th International Conference on Intelligence of Things 2026 (ICITconf'2026)},
     address = {Seoul, South Korea},
     days = {15},
     month = oct,
     year = {2026},
   }

2. APWeb-WAIM’26

   

   CorrTie: Correction-aware Tie-breaking for Active Learning with Vision-Language Models

   Duc Tai Phan, Tri Nguyen Thai, Khoa Minh Nguyen, Nhut Minh Nguyen, and Duc Ngoc Minh Dang

   In The 10th APWeb-WAIM Joint International Conference on Web and Big Data (APWeb-WAIM 2026) , Da Nang, Vietnam, Sep 2026

   Abs Bib Code

   Vision-language models (VLMs) provide strong transferable representations for visual recognition, but adapting them to downstream image classification tasks remains challenging when labeled data are scarce. Active learning (AL) reduces annotation cost by selecting informative unlabeled samples for oracle labeling; however, most AL strategies use oracle feedback only to expand the labeled training set. The correction event between a model prediction and the oracle label is usually not reused to guide future acquisition. To address this limitation, we propose Correction-aware Tie-breaking (CorrTie), a lightweight post-acquisition reranker for AL with VLMs. CorrTie keeps the base AL strategy unchanged and uses past oracle corrections to estimate class-wise correction risk. This risk is then used to rerank near-tie candidates while preserving the original acquisition ranking as the primary signal. Experiments with CoOp prompt learning on four image classification benchmarks and five base AL strategies show that CorrTie improves the average accuracy by +7.14, +3.74, and +0.66 percentage points at the 1%, 2%, and 5% annotation budgets, respectively. These results indicate that oracle corrections can serve not only as supervision for prompt learning, but also as useful feedback for future active sample acquisition.

   @inproceedings{Phan2607CorrTie,
     author = {Phan, Duc Tai and Thai, Tri Nguyen and Nguyen, Khoa Minh and Nguyen, Nhut Minh and Dang, Duc Ngoc Minh},
     title = {CorrTie: Correction-aware Tie-breaking for Active Learning with Vision-Language Models},
     booktitle = {The 10th APWeb-WAIM Joint International Conference on Web and Big Data (APWeb-WAIM 2026)},
     address = {Da Nang, Vietnam},
     days = {7},
     month = sep,
     year = {2026},
   }

3. KBS

   

   From object difficulty to image scoring: A strategy for active learning in object detection

   Duc Tai Phan, Nhut Minh Nguyen, Khang Phuc Nguyen, Phuong-Nam Tran, Nhat Truong Pham, Linh Le, Choong Seon Hong, and Duc Ngoc Minh Dang

   Knowledge-Based Systems, Apr 2026

   Abs Bib HTML Code

   Object detection (OD) faces the costly hurdle of large-scale dataset annotation. Active Learning (AL) is a promising solution that selects the most beneficial samples for annotation. Applying AL to OD is challenging, as it requires addressing uncertainties in both classifying and localizing objects while combining object-level information into a single image-level decision. Current methods, like uncertainty sampling, often rely on output-level uncertainty or heuristic signal combinations. However, relying on static model outputs limits their ability to capture underlying feature-space instabilities. To address these limitations, we propose Feature Difficulty-based AL (FDAL), a novel framework that shifts the focus from output-level heuristics to active latent-space manipulation. By systematically interpolating unlabeled features toward class anchors, FDAL identifies prediction inconsistencies to reveal hidden traits unknown to the model. This anchor-guided interpolation unifies classification and localization uncertainties, capturing latent instabilities that traditional paradigms miss. Experiments on four OD benchmarks demonstrate that FDAL achieves state-of-the-art (SOTA) results, consistently improving detection accuracy while substantially reducing annotation costs. FDAL outperforms SOTA methods by 0.8% in detection accuracy on Pattern analysis, statistical modelling, and computational learning (PASCAL) Visual Object Classes (PASCAL VOC), 1.99% on Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI), 1.68% on Cityscapes, and 0.27% on Microsoft Common Objects in Context (MS COCO). FDAL’s selection is highly efficient, taking 0.06 seconds per image on PASCAL VOC and up to 0.63 seconds per image on Cityscapes, enabling scalability to large unlabeled pools. FDAL offers a practical, efficient, and effective solution for advancing AL in OD.

   @article{PHAN2026115946,
     author = {Phan, Duc Tai and Nguyen, Nhut Minh and Nguyen, Khang Phuc and Tran, Phuong-Nam and Pham, Nhat Truong and Le, Linh and Hong, Choong Seon and Dang, Duc Ngoc Minh},
     title = {From object difficulty to image scoring: A strategy for active learning in object detection},
     journal = {Knowledge-Based Systems},
     month = apr,
     year = {2026},
     keywords = {Active learning, Object detection, Computer vision, Difficulty, Instability},
     volume = {342},
     pages = {115946},
     issn = {0950-7051},
     doi = {https://doi.org/10.1016/j.knosys.2026.115946},
     url = {https://www.sciencedirect.com/science/article/pii/S0950705126006726}
   }

2025

1. MEDES’25

   

   DAAL: Dual Ambiguity in Active Learning for Object Detection with YOLOE

   Duc Tai Phan, Nhut Minh Nguyen, and Duc Ngoc Minh Dang

   In 17th International Conference on Management of Digital Ecosystems , Ho Chi Minh City, Vietnam, Nov 2025

   Abs Code

   The high cost of annotating vast datasets poses a significant challenge in object detection, hindering the development of robust models. By selectively annotating the most instructive samples, active learning has become a viable strategy for addressing this problem and improving model performance and labeling efficiency. However, the two main types of model uncertainty—epistemic, which captures the model’s uncertainty in classifying an object, and aleatoric, which addresses the inherent ambiguity in an object’s presence and location—are frequently not balanced in traditional active learning techniques. In this paper, we introduce DAAL (Dual Ambiguity Active Learning), a novel framework that quantifies and combines both epistemic and aleatoric ambiguity into a single, weighted score. Epistemic ambiguity measures the model’s indecision in assigning semantic labels, while aleatoric ambiguity assesses the conviction in object presence and localization. By combining these, DAAL selects the most informative images for annotation, optimizing model performance under limited labeling budgets. Extensive experiments on popular benchmarks demonstrate that DAAL consistently outperforms traditional methods, achieving superior accuracy under the same limited labeling budget. This affirms its effectiveness in creating more efficient annotation workflows for object detection.

2. APNOMS’25

   

   ALMUS: Enhancing Active Learning for Object Detection with Metric-Based Uncertainty Sampling

   Duc Tai Phan, Nhut Minh Nguyen, Khang Phuc Nguyen, Tri Minh Pham, and Duc Ngoc Minh Dang

   In 2025 25th Asia-Pacific Network Operations and Management Symposium (APNOMS’2025) , Kaohsiung, Taiwan, Sep 2025

   Abs Bib HTML

   Object detection is critical in computer vision but often requires large amounts of labeled data for effective training. Active learning (AL) has emerged as a promising solution to reduce the annotation burden by selecting the most informative samples for labeling. However, existing AL methods for object detection primarily focus on uncertainty sampling, which may not effectively balance the dual challenges of classification and localization. In this study, we explore active learning for object detection, with the objective of optimizing model performance while substantially reducing the demand for annotated data. We propose a novel Active Learning with Metric-based Uncertainty Sampling (ALMUS) that works effectively for the object detection task. This approach prioritizes selecting images containing objects from categories where the model exhibits suboptimal performance, as determined by category-specific evaluation metrics. To balance the annotation budget across different object classes, we propose a dynamic allocation strategy that considers the difficulty of each class and the distribution of object instances within the dataset. This combination of strategies enables our method to effectively address the dual challenges of classification and localization in object detection tasks while still focusing on the rarest and most challenging classes. We conduct extensive experiments on the PASCAL VOC 2007 and 2012 datasets, demonstrating that our method outperforms several active learning baselines. Our results indicate that the proposed approach enhances model performance and accelerates convergence, making it a valuable contribution to the field of active learning in object detection.

   @inproceedings{11181447,
     author = {Phan, Duc Tai and Nguyen, Nhut Minh and Nguyen, Khang Phuc and Pham, Tri Minh and Dang, Duc Ngoc Minh},
     title = {ALMUS: Enhancing Active Learning for Object Detection with Metric-Based Uncertainty Sampling},
     booktitle = {2025 25th Asia-Pacific Network Operations and Management Symposium (APNOMS'2025)},
     address = {Kaohsiung, Taiwan},
     pages = {6},
     month = sep,
     year = {2025},
     keywords = {Location awareness;Measurement;Training;Uncertainty;Annotations;Active learning;Object detection;Robustness;Resource management;Convergence;Active Learning;Object Detection;Uncertainty Sampling;Metric-based Sampling},
     doi = {10.23919/APNOMS67058.2025.11181447},
     isbn = {978-1-23456-789-0},
   }