Tanima Dutta
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
Tanima Dutta has a total of 51 co-authors. Here are the top co-authors by number of shared publications:
- Hari Prabhat Gupta
Department of Computer Science and Engineering, Institute of Technology (BHU) Varanasi, Varanasi, India
28 shared publications - Rahul Mishra
Indian Institute of Technology Patna
10 shared publications - Deepali Verma
Indian Institute of Technology (BHU), Varanasi, India
8 shared publications Hari Prabhat Gupta
IIT (BHU) Varanasi
6 shared publications- Preti Kumari
School of Computing and Electrical Engineering, IIT Mandi, Mandi, India
6 shared publications - Ashish Gupta
BITS Pilani Dubai Campus
5 shared publications
Latest Publications
- 29 May 2026
- Published by Elsevier in Computers and Electrical Engineering
- Vol. 137
SWD-DA: Sliced Wasserstein Distance for manifold-aware domain adaptation
- 29 May 2026
- Published by Elsevier in Computers and Electrical Engineering
- Vol. 137
SWD-DA: Sliced Wasserstein Distance for manifold-aware domain adaptation
- 23 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE MultiMedia
- Vol. PP (1070986X) , 1-11
Behaviour Modelling Aware Fine-Grained Video Captioning
In video captioning, current methods often struggle with generating accurate descriptions due to challenges in modeling object interactions, especially predicates that rely on both object dynamics and motion patterns. To address this, we introduce the Behavior Modeling-Aware Video Captioning Network (BMVCap), which enhances video captioning by capturing fine-grained details of interactions, co-occurring objects, and contextual background using a transformer-based encoder for each modality stream. The model integrates these streams through an adaptive fusion mechanism in the transformer decoder, allowing it to generate more precise captions. Additionally, BMVCap employs a caption length control mechanism and optimized reinforcement learning, maximizing rewards from multiple evaluation metrics. Extensive experiments on the MSVD, MSR-VTT, and VATEX datasets show that BMVCap significantly improves captioning accuracy by better modeling complex interactions and activity attributes. The results demonstrate that emphasizing complex interactions and activity attributes leads to substantial improvements in the accuracy and reliability of video captioning.
- 23 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE MultiMedia
- Vol. PP (1070986X) , 1-11
Behaviour Modelling Aware Fine-Grained Video Captioning
In video captioning, current methods often struggle with generating accurate descriptions due to challenges in modeling object interactions, especially predicates that rely on both object dynamics and motion patterns. To address this, we introduce the Behavior Modeling-Aware Video Captioning Network (BMVCap), which enhances video captioning by capturing fine-grained details of interactions, co-occurring objects, and contextual background using a transformer-based encoder for each modality stream. The model integrates these streams through an adaptive fusion mechanism in the transformer decoder, allowing it to generate more precise captions. Additionally, BMVCap employs a caption length control mechanism and optimized reinforcement learning, maximizing rewards from multiple evaluation metrics. Extensive experiments on the MSVD, MSR-VTT, and VATEX datasets show that BMVCap significantly improves captioning accuracy by better modeling complex interactions and activity attributes. The results demonstrate that emphasizing complex interactions and activity attributes leads to substantial improvements in the accuracy and reliability of video captioning.
- 16 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 33 (10709908) , 1155-1159
Confidence-Aware Optimal Transport for Open Set Time Series Adaptation Under Non-Stationary Shifts
Open Set Domain Adaptation (OSDA) in non-stationary time series remains a challenging problem, particularly when the target domain contains previously unseen classes and exhibits temporal drift due to changing behaviors, sensor noise, or environmental variability. Existing adaptation methods often fail in such settings, misaligning unknown target samples and resulting in negative transfer. We propose a Confidence-Aware Latent Optimal Transport framework, named as CALOT, that performs selective domain alignment in a learned latent space. CALOT filters out low-confidence target samples—those likely to belong to unknown classes—using entropy-based confidence scoring, and applies optimal transport only to the most reliable instances. A self-regularization loss further enhances alignment stability under temporal variability. Our work achieves 90% accuracy on known classes and 80% accuracy in rejecting unknown classes, resulting in a strong H-score of 85% on the challenging non-stationary HHAR dataset. Additionally, it obtains H-scores of 76% on WISDM, 73% on HAR, 74% on EDF, and 75% on the BOILER dataset.
- 16 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 33 (10709908) , 1155-1159
Confidence-Aware Optimal Transport for Open Set Time Series Adaptation Under Non-Stationary Shifts
Open Set Domain Adaptation (OSDA) in non-stationary time series remains a challenging problem, particularly when the target domain contains previously unseen classes and exhibits temporal drift due to changing behaviors, sensor noise, or environmental variability. Existing adaptation methods often fail in such settings, misaligning unknown target samples and resulting in negative transfer. We propose a Confidence-Aware Latent Optimal Transport framework, named as CALOT, that performs selective domain alignment in a learned latent space. CALOT filters out low-confidence target samples—those likely to belong to unknown classes—using entropy-based confidence scoring, and applies optimal transport only to the most reliable instances. A self-regularization loss further enhances alignment stability under temporal variability. Our work achieves 90% accuracy on known classes and 80% accuracy in rejecting unknown classes, resulting in a strong H-score of 85% on the challenging non-stationary HHAR dataset. Additionally, it obtains H-scores of 76% on WISDM, 73% on HAR, 74% on EDF, and 75% on the BOILER dataset.
- 23 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE MultiMedia
- Vol. PP (1070986X) , 1-13
Privacy-Preserving Out-of-Distribution Alignment for Federated GNNs
Out-of-distribution (OOD) generalization is a major challenge in federated graph learning, where client data distributions often differ significantly due to shifts in node features and structural patterns. This non-IID nature leads to poor performance of Graph Neural Networks (GNNs) trained solely on federated task losses. While domain-alignment techniques—such as adversarial training, invariant risk minimization, and kernel-based methods like Maximum Mean Discrepancy (MMD)—offer promising solutions for mitigating distribution shifts, they typically require access to shared embeddings, which compromises user privacy. We propose FedMMD-P, a framework for privacy-preserving alignment in federated out-of-distribution graph learning. Each client computes a compact, noisy RFF-based sketch of its node-embedding distribution, enabling a DP-MMD penalty with only O( nD) local cost and D-dimensional communication per round. Across five benchmarks (Reddit, OGB-Arxiv, Ethereum, Tox21, PubMed), FedMMD-P delivers up to 6 percentage-point accuracy gains over non-private baselines under severe domain shifts while incurring <0.1% communication overhead and only 5–10% runtime overhead. We derive end-to-end (ε, δ)-DP bounds via Moments Accountant and Rényi composition, demonstrating practical privacy guarantees.
- 23 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE MultiMedia
- Vol. PP (1070986X) , 1-13
Privacy-Preserving Out-of-Distribution Alignment for Federated GNNs
Out-of-distribution (OOD) generalization is a major challenge in federated graph learning, where client data distributions often differ significantly due to shifts in node features and structural patterns. This non-IID nature leads to poor performance of Graph Neural Networks (GNNs) trained solely on federated task losses. While domain-alignment techniques—such as adversarial training, invariant risk minimization, and kernel-based methods like Maximum Mean Discrepancy (MMD)—offer promising solutions for mitigating distribution shifts, they typically require access to shared embeddings, which compromises user privacy. We propose FedMMD-P, a framework for privacy-preserving alignment in federated out-of-distribution graph learning. Each client computes a compact, noisy RFF-based sketch of its node-embedding distribution, enabling a DP-MMD penalty with only O( nD) local cost and D-dimensional communication per round. Across five benchmarks (Reddit, OGB-Arxiv, Ethereum, Tox21, PubMed), FedMMD-P delivers up to 6 percentage-point accuracy gains over non-private baselines under severe domain shifts while incurring <0.1% communication overhead and only 5–10% runtime overhead. We derive end-to-end (ε, δ)-DP bounds via Moments Accountant and Rényi composition, demonstrating practical privacy guarantees.
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE MultiMedia
- Vol. 33 (1) , 7-20
Toward Improving Arbitrary-Shaped Text Detection With Boundary Adaptation in Noisy Scene Images
Accurate detection of arbitrary-shaped text in natural scenes remains a challenging problem, particularly under real-world noise conditions such as blur, occlusion, uneven lighting, and cluttered backgrounds. These factors degrade local features and boundary precision, hindering the performance of existing contour-based methods. To address this, we propose Ray Contour Net (RC-Net), a robust framework for boundary adaptation in noisy environments. RC-Net incorporates a Noise- Aware Attention Module to obtain noise-robust features and an Active Contour Point Initialization Module that emits contour points in a polar coordinate system anchored at local text centers. These are refined by a Deformation Module using lightweight transformers to ensure shape-aware, noise-tolerant adaptation. Additionally, we introduce a Tightness Evaluation Loss (TEL) to enforce precise boundary predictions. Extensive experiments on six benchmark datasets including Total-Text, CTW-1500, MSRATD500, ICDAR 2015, ICDAR2019 ArT, and ICDAR2019 MLT demonstrate consistent F-measure improvements of 2% to 3% over strong baselines, validating the effectiveness and robustness of our method.
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE MultiMedia
- Vol. 33 (1) , 7-20
Toward Improving Arbitrary-Shaped Text Detection With Boundary Adaptation in Noisy Scene Images
Accurate detection of arbitrary-shaped text in natural scenes remains a challenging problem, particularly under real-world noise conditions such as blur, occlusion, uneven lighting, and cluttered backgrounds. These factors degrade local features and boundary precision, hindering the performance of existing contour-based methods. To address this, we propose Ray Contour Net (RC-Net), a robust framework for boundary adaptation in noisy environments. RC-Net incorporates a Noise- Aware Attention Module to obtain noise-robust features and an Active Contour Point Initialization Module that emits contour points in a polar coordinate system anchored at local text centers. These are refined by a Deformation Module using lightweight transformers to ensure shape-aware, noise-tolerant adaptation. Additionally, we introduce a Tightness Evaluation Loss (TEL) to enforce precise boundary predictions. Extensive experiments on six benchmark datasets including Total-Text, CTW-1500, MSRATD500, ICDAR 2015, ICDAR2019 ArT, and ICDAR2019 MLT demonstrate consistent F-measure improvements of 2% to 3% over strong baselines, validating the effectiveness and robustness of our method.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 174313-174337
From Generic to News Image Captioning: A Survey of Entity-Centric Multimodal Models
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 174313-174337
Open AccessFrom Generic to News Image Captioning: A Survey of Entity-Centric Multimodal Models
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. 13 (5) , 1-13
Meta-Adaptive Continual Learning for Dynamic Social Communication Graphs
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. 13 (5) , 1-13
Meta-Adaptive Continual Learning for Dynamic Social Communication Graphs
- 2 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. 12 (6) , 4784-4796
Enhancing User Engagement Through Contextual Video Captioning With Social and Behavior Insights
In the era of social media, accurate and contextually rich captions are essential for enhancing user engagement with video content. Traditional captioning methods, relying solely on visual content, often fall short of capturing audience interest. To address this, advanced captioning solutions are needed to combine visual cues with contextual insights. This article proposes a novel captioning approach that incorporates contextual information, leading to more accurate and engaging captions. By extracting human behavior, object interactions, and social cues, the method generates relevant and informative captions. To further refine the learning process, metalearning is employed to optimize a reward function. The effectiveness of the proposed method is validated on Microsoft video description (MSVD), Microsoft research-video to text (MSR-VTT), and video and text (VATEX) benchmark datasets through ablation studies and detailed experiments. The results demonstrate that contextual video captioning significantly enhances user experience. Behavioral features enrich feature representations, and context learning helps in predicting accurate and informative captions.
- 2 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. 12 (6) , 4784-4796
Enhancing User Engagement Through Contextual Video Captioning With Social and Behavior Insights
In the era of social media, accurate and contextually rich captions are essential for enhancing user engagement with video content. Traditional captioning methods, relying solely on visual content, often fall short of capturing audience interest. To address this, advanced captioning solutions are needed to combine visual cues with contextual insights. This article proposes a novel captioning approach that incorporates contextual information, leading to more accurate and engaging captions. By extracting human behavior, object interactions, and social cues, the method generates relevant and informative captions. To further refine the learning process, metalearning is employed to optimize a reward function. The effectiveness of the proposed method is validated on Microsoft video description (MSVD), Microsoft research-video to text (MSR-VTT), and video and text (VATEX) benchmark datasets through ablation studies and detailed experiments. The results demonstrate that contextual video captioning significantly enhances user experience. Behavioral features enrich feature representations, and context learning helps in predicting accurate and informative captions.
- 9 November 2025
- Published by Springer Nature
- p. 466-481
Mitigating Oversmoothing in GNNs Across Diverse Graph Domains
- 9 November 2025
- Published by Springer Nature
- p. 466-481
Mitigating Oversmoothing in GNNs Across Diverse Graph Domains