Abstract: The Segment Anything Model (SAM) achieves strong zero-shot segmentation, but adapting it to recognize a user-specified object across images without manual prompting or retraining remains an open problem. Existing training-free methods extract a target embedding from a reference image and use feature similarity to generate automatic prompts. However, we find that both raw SAM and DINOv2 features produce unreliable localization in practice. The root cause is grid artifacts: periodic spatial distortions from patch tokenization that corrupt similarity maps and displace prompts onto background regions. To alleviate this, we introduce Dual-TCR, a framework that explicitly decouples semantic localization from geometric segmentation. Specifically, we construct spatially coherent similarity maps via NVIDIA RADIO to mitigate patch grid artifacts, providing reliable prompts that guide SAM purely as a geometric mask decoder. To autonomously resolve the ambiguity between competing mask candidates, we propose the Target Consistency Reward (TCR), a zeroshot arbitrator scoring each candidate via feature alignment with the reference target embedding. We further introduce a Hybrid Refining Module (HRM) step, fusing outputs from both branches to produce a refined final output mask. Evaluated on the PerSeg benchmark, Dual-TCR achieves93.42% mIoU and 75.60% bIoU, establishing a new state-of-the-art specifically within the trainingfree category.

Citation: Minh-Khanh Le, Thanh-Danh Nguyen†, Vinh-Tiep Nguyen, and Tam V. Nguyen. “Dual-TCR: A Dual-Branch Target Consistency Reward Framework for Training-Free Personalized Segmentation.” Pattern Recognition Letters, Oct 2026.

Abstract: Skin lesion segmentation is critical for early melanoma detection, yet the high computational overhead of generative diffusion-based models hinders practical clinical deployment. While knowledge distillation (KD) offers a promising approach toward model compression, existing spatial-domain methods often fail to capture the unique coarse-to-fine denoising dynamics inherent in diffusion processes. To address this limitation, we propose Temporal Frequency Knowledge Distillation (TFKD), a novel timestep-aware framework that operates in the Fourier domain to dynamically align the spectral characteristics of the teacher and student models across the denoising trajectory. By explicitly transferring structural and boundary knowledge according to the spectral progression of different denoising stages, TFKD ensures a lightweight student model robustly retains the generative prior of its larger counterpart. Extensive evaluations on the ISIC2018 and HAM10000 benchmarks demonstrate that TFKD consistently outperforms existing distillation techniques, achieving a superior balance between segmentation accuracy and inference efficiency. Furthermore, generalization experiments on the DUTS dataset reveal that the distilled model remains competitive beyond dermoscopic imaging. These results establish TFKD as a simple, highly effective, and plug-and-play strategy for deploying compact diffusion-based segmentation models in both medical and general vision tasks.

Citation: Viet-Hoang Doan, Thanh-Danh Nguyen†, Vinh-Tiep Nguyen, and Tam V. Nguyen. “Knowledge Distillation via Temporal Frequency for Efficient Generative Semantic Image Segmentation.” 20th International Conference on Advanced Computing and Analytics. IEEE, 2026.

Abstract: Camouflage object detection and instance segmentation remain a challenging frontier in computer vision due to the intrinsic high similarity between foreground objects and their background surroundings. Furthermore, the scarcity of annotated camouflage data exacerbates the difficulty of training robust models, particularly in data-sparse regimes. In this paper, we introduce CAMO-InstSynth, a novel data enhancement framework exploiting the background context understanding to address few-shot camouflage object detection and instance segmentation. Accordingly, we propose a Multi-Conditional Background Synthesis Module that utilizes diffusion models to generate diverse, highfidelity backgrounds that maintain semantic consistency with camouflaged foregrounds. Unlike traditional augmentation techniques, which increase only the instance diversity, our method further conditions the synthesis process to simulate the camouflage environment. We validate our approach on the common CAMO-FS dataset over strong few-shot baselines. Our experiments demonstrate that CAMO-InstSynth significantly outperforms state-of-the-art methods, improving the iFS-RCNN baseline in both Segmentation and Detection tasks.

Citation: Thanh-Danh Nguyen, Vinh-Tiep Nguyen†, Kunpeng Li, and Tam V. Nguyen. “CAMO-InstSynth: Few-shot Camouflage Instance Segmentation with Multi-Conditional Background Synthesis and Generative Augmentation.” 18th Asian Conference on Intelligent Information and Database Systems, 2026.

Abstract: Accurate skin lesion segmentation is crucial for early melanoma detection and reliable medical diagnosis. Recently, such generative diffusion-based segmentation models have achieved state-of-the-art performance in this task. However, their large number of parameters hinders the deployment in real-world clinical and mobile settings. A key challenge is how to compress these models while preserving their reliability in diagnosis. To this end, we propose an edge-focused knowledge distillation (EFKD) strategy for the generative diffusion-based skin lesion segmentation, prioritizing contour information transfer from teacher to student. Accordingly, we introduce LiteDermoSegDiff, a lightweight diffusion-based lesion segmentation model that integrates knowledge distillation with boundary-aware supervision. This design ensures that the student network maintains sharp lesion boundaries while significantly reducing model parameters. Extensive experiments on the common ISIC2018 and HAM10000 dermoscopic benchmarks demonstrate that our method achieves up to 88.9% parameter reduction while maintaining competitive segmentation accuracy. These results show that diffusion-based segmentation models can be substantially compressed without compromising boundary awareness, paving the way for efficient and clinically deployable solutions in resource-constrained environments.

Citation: Viet-Hoang Doan, Thanh-Danh Nguyen†, Tuan-Kiet Ngo, and Vinh-Tiep Nguyen. “A Knowledge Distillation Approach for Diffusion-based Generative Skin Lesion Segmentation.” 2025 RIVF International Conference on Computing and Communication Technologies. IEEE, 2025.

Abstract: Identifying camouflaged instances is a critical yet underexplored problem in computer vision, where traditional segmentation models often fail due to extreme visual similarity between foreground and background. While recent advances have shown promise with deep learning models, they heavily depend on large annotated datasets, which are costly and impractical to collect in camouflage scenarios. In this work, we tackle this limitation by introducing a novel framework, dubbed CAMOGenOS, that leverages one-shot annotated samples to drive a generative process for data enrichment. Our approach integrates prompt-guided and mask-conditioned generative mechanisms to synthesize diverse, high-fidelity camouflaged instances, thereby enhancing the learning capacity of segmentation models under minimal supervision. We demonstrate the effectiveness of our CAMO-GenOS by setting up a novel state-of-the-art baseline for one-shot camouflage instance segmentation research on the challenging CAMO-FS benchmark.

Citation: Thanh-Danh Nguyen, Vinh-Tiep Nguyen†, and Tam V. Nguyen. “Generative One-shot Camouflage Instance Segmentation.” 2025 International Conference on Multimedia Analysis and Pattern Recognition. IEEE, 2025.

Abstract: Few-shot instance segmentation is an intense yet essential task, particularly in camouflaged scenarios where visual ambiguity between foreground and background makes instancelevel recognition more difficult. Prior approaches primarily focused on image augmentations in the color space domain to provide diverse perspective information to the segmentation models. However, this type of augmentation often fails to capture the full range of visual characteristics needed for robust generalization, particularly in camouflage images, due to the limited similar representation in the color space domain. To this end, we tackle this gap by exploiting a novel approach to augment and enhance image features in the derivative frequency domain. Accordingly, we propose a novel framework tailored for few-shot camouflage instance segmentation via the instance-aware frequency-based augmentation, dubbed FS-CAMOFreq, to enhance image diversity while preserving semantic structure, thereby improving the ability of the few-shot segmentor to learn from limited data. Extensive experiments on the challenging CAMO-FS benchmark demonstrate that our approach achieves superior performance compared to state-of-the-art baselines.

Citation: Thanh-Danh Nguyen, Hung-Phu Cao, Thanh Duc Ngo, Vinh-Tiep Nguyen†, and Tam V. Nguyen. “Few-Shot Instance Segmentation: An Exploration in the Frequency Domain for Camouflage Instances.” 2025 International Conference on Multimedia Analysis and Pattern Recognition. IEEE, 2025.

Abstract: Instance image segmentation task requires training with abundant annotated data to achieve high accuracy. Recently, conditional image synthesis has demonstrated its effectiveness in generating synthetic data for this task. However, existing image synthesis models face challenges in generating target instances to match the masks with complex shapes. Moreover, others fail to create diverse instances due to utilizing low-context simple text prompts. To address these issues, we propose CAMUL, a framework for context-aware multiconditional instance synthesis. CAMUL introduces two key innovations: CARP (cross-attention refinement prompting) to enhance the alignment of generated instances with conditional masks, and iCAFF (incremental context-aware feature fusion) to determine the general embeddings of the instances for a more precise context understanding. Our method significantly improves segmentation performance, increasing up to 15.34% average precision (AP) on Cityscapes and 3.34% AP on the large-scale ADE20K benchmark compared to the baselines.

Citation: Thanh-Danh Nguyen, Trong-Tai Dam Vu, Bich-Nga Pham, Thanh Duc Ngo, Tam V. Nguyen, and Vinh-Tiep Nguyen†. “CAMUL: Context-Aware Multi-conditional Instance Synthesis for Image Segmentation.” IEEE MultiMedia, Jun 2025.

Abstract: Semantic segmentation plays a crucial role in traffic scene understanding, especially in nighttime conditions. This paper tackles the task of semantic segmentation in nighttime scenes. The largest challenge of this task is the lack of annotated nighttime images to train a deep learning-based scene parser. The existing annotated datasets are abundant in daytime conditions but scarce in nighttime due to the high cost. Thus, we propose a novel Label Transfer Scene Parser (LTSP) framework for nighttime scene semantic segmentation by leveraging daytime annotation transfer. Our framework performs segmentation in the dark without training on real nighttime an-notated data. In particular, we propose translating daytime images to nighttime conditions to obtain more data with annotation in an efficient way. In addition, we utilize the pseudo-labels inferred from unlabeled nighttime scenes to further train the scene parser. The novelty of our work is the ability to perform nighttime segmentation via daytime annotated labels and nighttime synthetic versions of the same set of images. The extensive exper-iments demonstrate the improvement and efficiency of our scene parser over the state-of-the-art methods with a similar semi-supervised approach on the benchmark of Nighttime Driving Test dataset. Notably, our proposed method utilizes only one-tenth of the amount of labeled and unlabeled data in comparison with the previous methods.

Citation: Thanh-Danh Nguyen, Nguyen Phan, Tam V. Nguyen†, Vinh-Tiep Nguyen, and Minh-Triet Tran. “Nighttime Scene Understanding with Label Transfer Scene Parser.” Image and Vision Computing, Sep 2024.

Abstract: Camouflaged object detection and segmentation is a new and challenging research topic in computer vision. There is a serious issue of lacking data on concealed objects such as camouflaged animals in natural scenes. In this paper, we address the problem of few-shot learning for camouflaged object detection and segmentation. To this end, we first collect a new dataset, CAMO-FS, for the benchmark. As camouflaged instances are challenging to recognize due to their similarity compared to the surroundings, we guide our models to obtain camouflaged features that highly distinguish the instances from the background. In this work, we propose FS-CDIS, a framework to efficiently detect and segment camouflaged instances via two loss functions contributing to the training process. Firstly, the instance triplet loss with the characteristic of differentiating the anchor, which is the mean of all camouflaged foreground points, and the background points are employed to work at the instance level. Secondly, to consolidate the generalization at the class level, we present instance memory storage with the scope of storing camouflaged features of the same category, allowing the model to capture further class-level information during the learning process. The extensive experiments demonstrated that our proposed method achieves state-of-the-art performance on the newly collected dataset.

Citation: Thanh-Danh Nguyen*, Anh-Khoa Nguyen Vu*, Nhat-Duy Nguyen*, et al. “The Art of Camouflage: Few-shot Learning for Animal Detection and Segmentation.” IEEE Access, Jul 2024.

Abstract: Scene understanding at the instance level is an essential task in computer vision to support modern Advanced Driver Assistance Systems. Solutions have been proposed with abundant annotated training data. However, the annotation at the instance level is high-cost due to huge manual efforts. In this work, we solve this problem by introducing InstSynth, an advanced framework leveraging instance-wise annotations as conditions to enrich the training data. Existing methods focused on semantic segmentation via using prompts to synthesize image-annotation pairs, facing an unrealistic manner. Our proposals utilize the strength of such large generative models to synthesize instance data with prompt-guided and maskbased mechanisms to boost the performance of the instancelevel scene understanding models. We empirically improve the performance of the latest instance segmentation architectures of FastInst and OneFormer by 14.49% and 11.59% AP, respectively, evaluated on the Cityscapes benchmark. Accordingly, we construct an instance-level synthesized dataset, dubbed ISCityscapes, with over a 4× larger number of instances in comparison with the vanilla Cityscapes.

Citation: Thanh-Danh Nguyen, Bich-Nga Pham, Trong-Tai Dam Vu, et al. “InstSynth: Instance-wise Prompt-guided Style Masked Conditional Data Synthesis for Scene Understanding.” 2024 MAPR. IEEE, 2024.

Abstract: Understanding camouflage images at instance level is such a challenging task in computer vision. Since the camouflage instances have their colors and textures similar to the background, the key to distinguish them in the images should rely on their contours. The contours seperate the instance from the background, thus recognizing these contours should break their camouflage mechanism. To this end, we address the problem of camouflage instance segmentation via the Contour Emphasis approach. We improve the ability of the segmentation models by enhancing the contours of the camouflaged instances. We propose the CE-OST framework which employs the well-known architecture of Transformer-based models in a one-stage manner to boost the performance of camouflaged instance segmentation. The extensive experiments prove our contributions over the state-of-the-art baselines on different benchmarks, i.e. CAMO++, COD10K and NC4K.

Citation: Thanh-Danh Nguyen, Duc-Tuan Luu, Vinh-Tiep Nguyen†, and Thanh Duc Ngo. “CE-OST: Contour Emphasis for One-Stage Transformer-based Camouflage Instance Segmentation.” 2023 MAPR. IEEE, 2023.